Publications

RSS feed available.

2019

  • R. Akrour, J. Pajarinen, G. Neumann, and J. Peters, “Projections for approximate policy iteration algorithms,” in Proceedings of the international conference on machine learning (icml), 2019, p. 181–190.
    [BibTeX] [Abstract] [Download PDF]

    Approximate policy iteration is a class of reinforcement learning (RL) algorithms where the policy is encoded using a function approximator and which has been especially prominent in RL with continuous action spaces. In this class of RL algorithms, ensuring increase of the policy return during policy update often requires to constrain the change in action distribution. Several approximations exist in the literature to solve this constrained policy update problem. In this paper, we propose to improve over such solutions by introducing a set of projections that transform the constrained problem into an unconstrained one which is then solved by standard gradient descent. Using these projections, we empirically demonstrate that our approach can improve the policy update solution and the control over exploration of existing approximate policy iteration algorithms.

    @inproceedings{lirolem36285,
    publisher = {Proceedings of Machine Learning Research},
    booktitle = {Proceedings of the International Conference on Machine Learning (ICML)},
    year = {2019},
    author = {R. Akrour and J. Pajarinen and Gerhard Neumann and J. Peters},
    title = {Projections for Approximate Policy Iteration Algorithms},
    pages = {181--190},
    month = {June},
    keywords = {ARRAY(0x558b53720790)},
    abstract = {Approximate policy iteration is a class of reinforcement learning (RL) algorithms where the policy is encoded using a function approximator and which has been especially prominent in RL with continuous action spaces. In this class of RL algorithms, ensuring increase of the policy return during policy update often requires to constrain the change in action distribution. Several approximations exist in the literature to solve this constrained policy update problem. In this paper, we propose to improve over such solutions by introducing a set of projections that transform the constrained problem into an unconstrained one which is then solved by standard gradient descent. Using these projections, we empirically demonstrate that our approach can improve the policy update solution and the control over exploration of existing approximate policy iteration algorithms.},
    url = {http://eprints.lincoln.ac.uk/36285/}
    }
  • P. Becker, H. Pandya, G. Gebhardt, C. Zhao, J. C. Taylor, and G. Neumann, “Recurrent kalman networks: factorized inference in high-dimensional deep feature spaces,” in Proceedings of the 36th international conference on machine learning, Long Beach, California, USA, 2019, p. 544–552.
    [BibTeX] [Abstract] [Download PDF]

    In order to integrate uncertainty estimates into deep time-series modelling, Kalman Filters (KFs) (Kalman et al., 1960) have been integrated with deep learning models, however, such approaches typically rely on approximate inference tech- niques such as variational inference which makes learning more complex and often less scalable due to approximation errors. We propose a new deep approach to Kalman filtering which can be learned directly in an end-to-end manner using backpropagation without additional approximations. Our approach uses a high-dimensional factorized latent state representation for which the Kalman updates simplify to scalar operations and thus avoids hard to backpropagate, computationally heavy and potentially unstable matrix inversions. Moreover, we use locally linear dynamic models to efficiently propagate the latent state to the next time step. The resulting network architecture, which we call Recurrent Kalman Network (RKN), can be used for any time-series data, similar to a LSTM (Hochreiter & Schmidhuber, 1997) but uses an explicit representation of uncertainty. As shown by our experiments, the RKN obtains much more accurate uncertainty estimates than an LSTM or Gated Recurrent Units (GRUs) (Cho et al., 2014) while also showing a slightly improved prediction performance and outperforms various recent generative models on an image imputation task.

    @inproceedings{lirolem36286,
    address = {Long Beach, California, USA},
    month = {June},
    pages = {544--552},
    series = {Proceedings of Machine Learning Research},
    title = {Recurrent Kalman Networks: Factorized Inference in High-Dimensional Deep Feature Spaces},
    booktitle = {Proceedings of the 36th International Conference on Machine Learning},
    volume = {97},
    publisher = {Proceedings of Machine Learning Research},
    author = {Philipp Becker and Harit Pandya and Gregor Gebhardt and Cheng Zhao and C. James Taylor and Gerhard Neumann},
    year = {2019},
    url = {http://eprints.lincoln.ac.uk/36286/},
    keywords = {ARRAY(0x558b537f8be8)},
    abstract = {In order to integrate uncertainty estimates into deep time-series modelling, Kalman Filters (KFs) (Kalman et al., 1960) have been integrated with deep learning models, however, such approaches typically rely on approximate inference tech- niques such as variational inference which makes learning more complex and often less scalable due to approximation errors. We propose a new deep approach to Kalman filtering which can be learned directly in an end-to-end manner using backpropagation without additional approximations. Our approach uses a high-dimensional factorized latent state representation for which the Kalman updates simplify to scalar operations and thus avoids hard to backpropagate, computationally heavy and potentially unstable matrix inversions. Moreover, we use locally linear dynamic models to efficiently propagate the latent state to the next time step. The resulting network architecture, which we call Recurrent Kalman Network (RKN), can be used for any time-series data, similar to a LSTM (Hochreiter \& Schmidhuber, 1997) but uses an explicit representation of uncertainty. As shown by our experiments, the RKN obtains much more accurate uncertainty estimates than an LSTM or Gated Recurrent Units (GRUs) (Cho et al., 2014) while also showing a slightly improved prediction performance and outperforms various recent generative models on an image imputation task.}
    }
  • P. Bosilj, I. Gould, T. Duckett, and G. Cielniak, “Pattern spectra from different component trees for estimating soil size distribution,” in 14th international symposium on mathematical morphology, 2019.
    [BibTeX] [Abstract] [Download PDF]

    We study the pattern spectra in context of soil structure analysis. Good soil structure is vital for sustainable crop growth. Accurate and fast measuring methods can contribute greatly to soil management decisions. However, the current in-field approaches contain a degree of subjectivity, while obtaining quantifiable results through laboratory techniques typically involves sieving the soil which is labour- and time-intensive. We aim to replace this physical sieving process through image analysis, and investigate the effectiveness of pattern spectra to capture the size distribution of the soil aggregates. We calculate the pattern spectra from partitioning hierarchies in addition to the traditional max-tree. The study is posed as an image retrieval problem, and confirms the ability of pattern spectra and suitability of different partitioning trees to re-identify soil samples in different arrangements and scales.

    @inproceedings{lirolem35548,
    booktitle = {14th International Symposium on Mathematical Morphology},
    journal = {International Symposium on Mathematical Morphology},
    publisher = {Springer},
    title = {Pattern Spectra from Different Component Trees for Estimating Soil Size Distribution},
    author = {Petra Bosilj and Iain Gould and Tom Duckett and Grzegorz Cielniak},
    year = {2019},
    month = {June},
    keywords = {ARRAY(0x558b53756aa0)},
    abstract = {We study the pattern spectra in context of soil structure analysis. Good soil structure is vital for sustainable crop growth. Accurate and fast measuring methods can contribute greatly to soil management decisions. However, the current in-field approaches contain a degree of subjectivity, while obtaining quantifiable results through laboratory techniques typically involves sieving the soil which is labour- and time-intensive. We aim to replace this physical sieving process through image analysis, and investigate the effectiveness of pattern spectra to capture the size distribution of the soil aggregates. We calculate the pattern spectra from partitioning hierarchies in addition to the traditional max-tree. The study is posed as an image retrieval problem, and confirms the ability of pattern spectra and suitability of different partitioning trees to re-identify soil samples in different arrangements and scales.},
    url = {http://eprints.lincoln.ac.uk/35548/}
    }
  • P. Bosilj, E. Aptoula, T. Duckett, and G. Cielniak, “Transfer learning between crop types for semantic segmentation of crops versus weeds in precision agriculture,” Journal of field robotics, 2019.
    [BibTeX] [Abstract] [Download PDF]

    Agricultural robots rely on semantic segmentation for distinguishing between crops and weeds in order to perform selective treatments, increase yield and crop health while reducing the amount of chemicals used. Deep learning approaches have recently achieved both excellent classification performance and real-time execution. However, these techniques also rely on a large amount of training data, requiring a substantial labelling effort, both of which are scarce in precision agriculture. Additional design efforts are required to achieve commercially viable performance levels under varying environmental conditions and crop growth stages. In this paper, we explore the role of knowledge transfer between deep-learning-based classifiers for different crop types, with the goal of reducing the retraining time and labelling efforts required for a new crop. We examine the classification performance on three datasets with different crop types and containing a variety of weeds, and compare the performance and retraining efforts required when using data labelled at pixel level with partially labelled data obtained through a less time-consuming procedure of annotating the segmentation output. We show that transfer learning between different crop types is possible, and reduces training times for up to \$80{$\backslash$}\%\$. Furthermore, we show that even when the data used for re-training is imperfectly annotated, the classification performance is within \$2{$\backslash$}\%\$ of that of networks trained with laboriously annotated pixel-precision data.

    @article{lirolem35535,
    journal = {Journal of Field Robotics},
    publisher = {Wiley Periodicals, Inc.},
    year = {2019},
    title = {Transfer learning between crop types for semantic segmentation of crops versus weeds in precision agriculture},
    author = {Petra Bosilj and Erchan Aptoula and Tom Duckett and Grzegorz Cielniak},
    abstract = {Agricultural robots rely on semantic segmentation for distinguishing between crops and weeds in order to perform selective treatments, increase yield and crop health while reducing the amount of chemicals used. Deep learning approaches have recently achieved both excellent classification performance and real-time execution. However, these techniques also rely on a large amount of training data, requiring a substantial labelling effort, both of which are scarce in precision agriculture. Additional design efforts are required to achieve commercially viable performance levels under varying environmental conditions and crop growth stages. In this paper, we explore the role of knowledge transfer between deep-learning-based classifiers for different crop types, with the goal of reducing the retraining time and labelling efforts required for a new crop. We examine the classification performance on three datasets with different crop types and containing a variety of weeds, and compare the performance and retraining efforts required when using data labelled at pixel level with partially labelled data obtained through a less time-consuming procedure of annotating the segmentation output. We show that transfer learning between different crop types is possible, and reduces training times for up to \$80{$\backslash$}\%\$. Furthermore, we show that even when the data used for re-training is imperfectly annotated, the classification performance is within \$2{$\backslash$}\%\$ of that of networks trained with laboriously annotated pixel-precision data.},
    keywords = {ARRAY(0x558b53597750)},
    url = {http://eprints.lincoln.ac.uk/35535/}
    }
  • F. Brandherm, J. Peters, G. Neumann, and R. Akrour, “Learning replanning policies with direct policy search,” Ieee robotics and automation letters (ra-l), vol. 4, iss. 2, p. 2196 –2203, 2019.
    [BibTeX] [Abstract] [Download PDF]

    Direct policy search has been successful in learning challenging real world robotic motor skills by learning open-loop movement primitives with high sample efficiency. These primitives can be generalized to different contexts with varying initial configurations and goals. Current state-of-the-art contextual policy search algorithms can however not adapt to changing, noisy context measurements. Yet, these are common characteristics of real world robotic tasks. Planning a trajectory ahead based on an inaccurate context that may change during the motion often results in poor accuracy, especially with highly dynamical tasks. To adapt to updated contexts, it is sensible to learn trajectory replanning strategies. We propose a framework to learn trajectory replanning policies via contextual policy search and demonstrate that they are safe for the robot, that they can be learned efficiently and that they outperform non-replanning policies for problems with partially observable or perturbed context

    @article{lirolem36284,
    pages = {2196 --2203},
    number = {2},
    month = {April},
    title = {Learning Replanning Policies with Direct Policy Search},
    volume = {4},
    year = {2019},
    author = {F. Brandherm and J. Peters and Gerhard Neumann and R. Akrour},
    journal = {IEEE Robotics and Automation Letters (RA-L)},
    url = {http://eprints.lincoln.ac.uk/36284/},
    keywords = {ARRAY(0x558b53805d10)},
    abstract = {Direct policy search has been successful in learning challenging real world robotic motor skills by learning open-loop movement primitives with high sample efficiency. These primitives can be generalized to different contexts with varying initial configurations and goals. Current state-of-the-art contextual policy search algorithms can however not adapt to changing, noisy context measurements. Yet, these are common characteristics of real world robotic tasks. Planning a trajectory ahead based on an inaccurate context that may change during the motion often results in poor accuracy, especially with highly dynamical tasks. To adapt to updated contexts, it is sensible to learn trajectory replanning strategies. We propose a framework to learn trajectory replanning policies via contextual policy search and demonstrate that they are safe for the robot, that they can be learned efficiently and that they outperform non-replanning policies for problems with partially observable or perturbed context}
    }
  • H. Cao, P. G. Esteban, M. Bartlett, P. Baxter, T. Belpaeme, E. Billing, H. Cai, M. Coeckelbergh, C. Costescu, D. David, A. D. Beir, D. Hernandez, J. Kennedy, H. Liu, S. Matu, A. Mazel, A. Pandey, K. Richardson, E. Senft, S. Thill, G. V. de Perre, B. Vanderborght, D. Vernon, K. Wakanuma, H. Yu, X. Zhou, and T. Ziemke, “Robot-enhanced therapy: development and validation of supervised autonomous robotic system for autism spectrum disorders therapy,” Ieee robotics & automation magazine, vol. 26, iss. 2, p. 49–58, 2019.
    [BibTeX] [Abstract] [Download PDF]

    Robot-assisted therapy (RAT) offers potential advantages for improving the social skills of children with autism spectrum disorders (ASDs). This article provides an overview of the developed technology and clinical results of the EC-FP7-funded Development of Robot-Enhanced therapy for children with AutisM spectrum disorders (DREAM) project, which aims to develop the next level of RAT in both clinical and technological perspectives, commonly referred to as robot-enhanced therapy (RET). Within this project, a supervised autonomous robotic system is collaboratively developed by an interdisciplinary consortium including psychotherapists, cognitive scientists, roboticists, computer scientists, and ethicists, which allows robot control to exceed classical remote control methods, e.g., Wizard of Oz (WoZ), while ensuring safe and ethical robot behavior. Rigorous clinical studies are conducted to validate the efficacy of RET. Current results indicate that RET can obtain an equivalent performance compared to that of human standard therapy for children with ASDs. We also discuss the next steps of developing RET robotic systems.

    @article{lirolem36203,
    title = {Robot-Enhanced Therapy: Development and Validation of Supervised Autonomous Robotic System for Autism Spectrum Disorders Therapy},
    volume = {26},
    publisher = {IEEE},
    month = {June},
    number = {2},
    pages = {49--58},
    author = {Hoang-Long Cao and Pablo G. Esteban and Madeleine Bartlett and Paul Baxter and Tony Belpaeme and Erik Billing and Haibin Cai and Mark Coeckelbergh and Cristina Costescu and Daniel David and Albert De Beir and Daniel Hernandez and James Kennedy and Honghai Liu and Silviu Matu and Alexandre Mazel and Amit Pandey and Kathleen Richardson and Emmanuel Senft and Serge Thill and Greet Van de Perre and Bram Vanderborght and David Vernon and Kutoma Wakanuma and Hui Yu and Xiaolong Zhou and Tom Ziemke},
    year = {2019},
    journal = {IEEE Robotics \& Automation Magazine},
    url = {http://eprints.lincoln.ac.uk/36203/},
    keywords = {ARRAY(0x558b5383a5e0)},
    abstract = {Robot-assisted therapy (RAT) offers potential advantages for improving the social skills of children with autism spectrum disorders (ASDs). This article provides an overview of the developed technology and clinical results of the EC-FP7-funded Development of Robot-Enhanced therapy for children with AutisM spectrum disorders (DREAM) project, which aims to develop the next level of RAT in both clinical and technological perspectives, commonly referred to as robot-enhanced therapy (RET). Within this project, a supervised autonomous robotic system is collaboratively developed by an interdisciplinary consortium including psychotherapists, cognitive scientists, roboticists, computer scientists, and ethicists, which allows robot control to exceed classical remote control methods, e.g., Wizard of Oz (WoZ), while ensuring safe and ethical robot behavior. Rigorous clinical studies are conducted to validate the efficacy of RET. Current results indicate that RET can obtain an equivalent performance compared to that of human standard therapy for children with ASDs. We also discuss the next steps of developing RET robotic systems.}
    }
  • C. Coppola, S. Cosar, D. R. Faria, and N. Bellotto, “Social activity recognition on continuous rgb-d video sequences,” International journal of social robotics, p. 1–15, 2019.
    [BibTeX] [Abstract] [Download PDF]

    Modern service robots are provided with one or more sensors, often including RGB-D cameras, to perceive objects and humans in the environment. This paper proposes a new system for the recognition of human social activities from a continuous stream of RGB-D data. Many of the works until now have succeeded in recognising activities from clipped videos in datasets, but for robotic applications it is important to be able to move to more realistic scenarios in which such activities are not manually selected. For this reason, it is useful to detect the time intervals when humans are performing social activities, the recognition of which can contribute to trigger human-robot interactions or to detect situations of potential danger. The main contributions of this research work include a novel system for the recognition of social activities from continuous RGB-D data, combining temporal segmentation and classification, as well as a model for learning the proximity-based priors of the social activities. A new public dataset with RGB-D videos of social and individual activities is also provided and used for evaluating the proposed solutions. The results show the good performance of the system in recognising social activities from continuous RGB-D data.

    @article{lirolem35151,
    pages = {1--15},
    publisher = {Springer},
    journal = {International Journal of Social Robotics},
    author = {Claudio Coppola and Serhan Cosar and Diego R. Faria and Nicola Bellotto},
    title = {Social Activity Recognition on Continuous RGB-D Video Sequences},
    year = {2019},
    abstract = {Modern service robots are provided with one or more sensors, often including RGB-D cameras, to perceive objects and humans in the environment. This paper proposes a new system for the recognition of human social activities from a continuous stream of RGB-D data. Many of the works until now have succeeded in recognising activities from clipped videos in datasets, but for robotic applications it is important to be able to move to more realistic scenarios in which such activities are not manually selected. For this reason, it is useful to detect the time intervals when humans are performing social activities, the recognition of which can contribute to trigger human-robot interactions or to detect situations of potential danger. The main contributions of this research work include a novel system for the recognition of social activities from continuous RGB-D data, combining temporal segmentation and classification, as well as a model for learning the proximity-based priors of the social activities. A new public dataset with RGB-D videos of social and individual activities is also provided and used for evaluating the proposed solutions. The results show the good performance of the system in recognising social activities from continuous RGB-D data.},
    keywords = {ARRAY(0x558b53717e10)},
    url = {http://eprints.lincoln.ac.uk/35151/}
    }
  • S. Cosar and N. Bellotto, “Human re-identification with a robot thermal camera using entropy-based sampling,” Journal of intelligent and robotic systems, 2019.
    [BibTeX] [Abstract] [Download PDF]

    Human re-identification is an important feature of domestic service robots, in particular for elderly monitoring and assistance, because it allows them to perform personalized tasks and human-robot interactions. However vision-based re-identification systems are subject to limitations due to human pose and poor lighting conditions. This paper presents a new re-identification method for service robots using thermal images. In robotic applications, as the number and size of thermal datasets is limited, it is hard to use approaches that require huge amount of training samples. We propose a re-identification system that can work using only a small amount of data. During training, we perform entropy-based sampling to obtain a thermal dictionary for each person. Then, a symbolic representation is produced by converting each video into sequences of dictionary elements. Finally, we train a classifier using this symbolic representation and geometric distribution within the new representation domain. The experiments are performed on a new thermal dataset for human re-identification, which includes various situations of human motion, poses and occlusion, and which is made publicly available for research purposes. The proposed approach has been tested on this dataset and its improvements over standard approaches have been demonstrated.

    @article{lirolem35778,
    author = {Serhan Cosar and Nicola Bellotto},
    title = {Human Re-Identification with a Robot Thermal Camera using Entropy-based Sampling},
    year = {2019},
    journal = {Journal of Intelligent and Robotic Systems},
    publisher = {Springer},
    keywords = {ARRAY(0x558b534acca8)},
    abstract = {Human re-identification is an important feature of domestic service robots, in particular for elderly monitoring and assistance, because it allows them to perform personalized tasks and human-robot interactions. However vision-based re-identification systems are subject to limitations due to human pose and poor lighting conditions. This paper presents a new re-identification method for service robots using thermal images. In robotic applications, as the number and size of thermal datasets is limited, it is hard to use approaches that require huge amount of training samples. We propose a re-identification system that can work using only a small amount of data. During training, we perform entropy-based sampling to obtain a thermal dictionary for each person. Then, a symbolic representation is produced by converting each video into sequences of dictionary elements. Finally, we train a classifier using this symbolic representation and geometric distribution within the new representation domain. The experiments are performed on a new thermal dataset for human re-identification, which includes various situations of human motion, poses and occlusion, and which is made publicly available for research purposes. The proposed approach has been tested on this dataset and its improvements over standard approaches have been demonstrated.},
    url = {http://eprints.lincoln.ac.uk/35778/}
    }
  • H. Cuayahuitl, “A data-efficient deep learning approach for deployable multimodal social robots,” Neurocomputing, 2019.
    [BibTeX] [Abstract] [Download PDF]

    The deep supervised and reinforcement learning paradigms (among others) have the potential to endow interactive multimodal social robots with the ability of acquiring skills autonomously. But it is still not very clear yet how they can be best deployed in real world applications. As a step in this direction, we propose a deep learning-based approach for efficiently training a humanoid robot to play multimodal games–-and use the game of `Noughts {$\backslash$}& Crosses’ with two variants as a case study. Its minimum requirements for learning to perceive and interact are based on a few hundred example images, a few example multimodal dialogues and physical demonstrations of robot manipulation, and automatic simulations. In addition, we propose novel algorithms for robust visual game tracking and for competitive policy learning with high winning rates, which substantially outperform DQN-based baselines. While an automatic evaluation shows evidence that the proposed approach can be easily extended to new games with competitive robot behaviours, a human evaluation with 130 humans playing with the \{{$\backslash$}it Pepper\} robot confirms that highly accurate visual perception is required for successful game play.

    @article{lirolem33533,
    author = {Heriberto Cuayahuitl},
    title = {A Data-Efficient Deep Learning Approach for Deployable Multimodal Social Robots},
    year = {2019},
    journal = {Neurocomputing},
    publisher = {Elsevier},
    note = {The final published version of this article can be accessed online at https://www.journals.elsevier.com/neurocomputing/},
    url = {http://eprints.lincoln.ac.uk/33533/},
    abstract = {The deep supervised and reinforcement learning paradigms (among others) have the potential to endow interactive multimodal social robots with the ability of acquiring skills autonomously. But it is still not very clear yet how they can be best deployed in real world applications. As a step in this direction, we propose a deep learning-based approach for efficiently training a humanoid robot to play multimodal games---and use the game of `Noughts {$\backslash$}\& Crosses' with two variants as a case study. Its minimum requirements for learning to perceive and interact are based on a few hundred example images, a few example multimodal dialogues and physical demonstrations of robot manipulation, and automatic simulations. In addition, we propose novel algorithms for robust visual game tracking and for competitive policy learning with high winning rates, which substantially outperform DQN-based baselines. While an automatic evaluation shows evidence that the proposed approach can be easily extended to new games with competitive robot behaviours, a human evaluation with 130 humans playing with the \{{$\backslash$}it Pepper\} robot confirms that highly accurate visual perception is required for successful game play.},
    keywords = {ARRAY(0x558b5356f768)}
    }
  • H. Cuayahuitl, D. Lee, S. Ryu, S. Choi, I. Hwang, and J. Kim, “Deep reinforcement learning for chatbots using clustered actions and human-likeness rewards,” in International joint conference on neural networks (ijcnn), 2019.
    [BibTeX] [Abstract] [Download PDF]

    Training chatbots using the reinforcement learning paradigm is challenging due to high-dimensional states, infinite action spaces and the difficulty in specifying the reward function. We address such problems using clustered actions instead of infinite actions, and a simple but promising reward function based on human-likeness scores derived from human-human dialogue data. We train Deep Reinforcement Learning (DRL) agents using chitchat data in raw text{–}without any manual annotations. Experimental results using different splits of training data report the following. First, that our agents learn reasonable policies in the environments they get familiarised with, but their performance drops substantially when they are exposed to a test set of unseen dialogues. Second, that the choice of sentence embedding size between 100 and 300 dimensions is not significantly different on test data. Third, that our proposed human-likeness rewards are reasonable for training chatbots as long as they use lengthy dialogue histories of ?10 sentences.

    @inproceedings{lirolem35954,
    month = {July},
    title = {Deep Reinforcement Learning for Chatbots Using Clustered Actions and Human-Likeness Rewards},
    author = {Heriberto Cuayahuitl and Donghyeon Lee and Seonghan Ryu and Sungja Choi and Inchul Hwang and Jihie Kim},
    year = {2019},
    booktitle = {International Joint Conference on Neural Networks (IJCNN)},
    publisher = {IEEE},
    url = {http://eprints.lincoln.ac.uk/35954/},
    keywords = {ARRAY(0x558b536a5b80)},
    abstract = {Training chatbots using the reinforcement learning paradigm is challenging due to high-dimensional states, infinite action spaces and the difficulty in specifying the reward function. We address such problems using clustered actions instead of infinite actions, and a simple but promising reward function based on human-likeness scores derived from human-human dialogue data. We train Deep Reinforcement Learning (DRL) agents using chitchat data in raw text{--}without any manual annotations. Experimental results using different splits of training data report the following. First, that our agents learn reasonable policies in the environments they get familiarised with, but their performance drops substantially when they are exposed to a test set of unseen dialogues. Second, that the choice of sentence embedding size between 100 and 300 dimensions is not significantly different on test data. Third, that our proposed human-likeness rewards are reasonable for training chatbots as long as they use lengthy dialogue histories of ?10 sentences.}
    }
  • K. Elgeneidy, P. Lightbody, S. Pearson, and G. Neumann, “Characterising 3d-printed soft fin ray robotic fingers with layer jamming capability for delicate grasping,” in Robosoft 2019, 2019.
    [BibTeX] [Abstract] [Download PDF]

    Motivated by the growing need within the agrifood industry to automate the handling of delicate produce, this paper presents soft robotic fingers utilising the Fin Ray effect to passively and gently adapt to delicate targets. The proposed Soft Fin Ray fingers feature thin ribs and are entirely 3D printed from a flexible material (NinjaFlex) to enhance their shape adaptation, compared to the original Fin Ray fingers. To overcome their reduced force generation, the effects of the angle and spacing of the flexible ribs were experimentally characterised. The results showed that at large displacements, layer jamming between tilted flexible ribs can significantly enhance the force generation, while minimal contact forces can be still maintained at small displacements for delicate grasping.

    @inproceedings{lirolem34950,
    title = {Characterising 3D-printed Soft Fin Ray Robotic Fingers with Layer Jamming Capability for Delicate Grasping},
    author = {Khaled Elgeneidy and Peter Lightbody and Simon Pearson and Gerhard Neumann},
    year = {2019},
    booktitle = {RoboSoft 2019},
    month = {June},
    url = {http://eprints.lincoln.ac.uk/34950/},
    abstract = {Motivated by the growing need within the agrifood industry to automate the handling of delicate produce, this paper presents soft robotic fingers utilising the Fin Ray effect to passively and gently adapt to delicate targets. The proposed Soft Fin Ray fingers feature thin ribs and are entirely 3D printed from a flexible material (NinjaFlex) to enhance their shape adaptation, compared to the original Fin Ray fingers. To overcome their reduced force generation, the effects of
    the angle and spacing of the flexible ribs were experimentally characterised. The results showed that at large displacements, layer jamming between tilted flexible ribs can significantly enhance the force generation, while minimal contact forces can be still maintained at small displacements for delicate grasping.},
    keywords = {ARRAY(0x558b5346dc68)}
    }
  • K. Elgeneidy, G. Neumann, S. Pearson, M. Jackson, and N. Lohse, “Contact detection and size estimation using a modular soft gripper with embedded flex sensors,” in International conference on intelligent robots and systems (iros 2018), 2019.
    [BibTeX] [Abstract] [Download PDF]

    Grippers made from soft elastomers are able to passively and gently adapt to their targets allowing deformable objects to be grasped safely without causing bruise or damage. However, it is difficult to regulate the contact forces due to the lack of contact feedback for such grippers. In this paper, a modular soft gripper is presented utilizing interchangeable soft pneumatic actuators with embedded flex sensors as fingers of the gripper. The fingers can be assembled in different configurations using 3D printed connectors. The paper investigates the potential of utilizing the simple sensory feedback from the flex and pressure sensors to make additional meaningful inferences regarding the contact state and grasped object size. We study the effect of the grasped object size and contact type on the combined feedback from the embedded flex sensors of opposing fingers. Our results show that a simple linear relationship exists between the grasped object size and the final flex sensor reading at fixed input conditions, despite the variation in object weight and contact type. Additionally, by simply monitoring the time series response from the flex sensor, contact can be detected by comparing the response to the known free-bending response at the same input conditions. Furthermore, by utilizing the measured internal pressure supplied to the soft fingers, it is possible to distinguish between power and pinch grasps, as the contact type affects the rate of change in the flex sensor readings against the internal pressure.

    @inproceedings{lirolem34713,
    author = {Khaled Elgeneidy and Gerhard Neumann and Simon Pearson and Michael Jackson and Niels Lohse},
    title = {Contact Detection and Size Estimation Using a Modular Soft Gripper with Embedded Flex Sensors},
    year = {2019},
    booktitle = {International Conference on Intelligent Robots and Systems (IROS 2018)},
    month = {January},
    abstract = {Grippers made from soft elastomers are able to passively and gently adapt to their targets allowing deformable objects to be grasped safely without causing bruise or damage. However, it is difficult to regulate the contact forces due to the lack of contact feedback for such grippers. In this paper, a modular soft gripper is presented utilizing interchangeable soft pneumatic actuators with embedded flex sensors as fingers of the gripper. The fingers can be assembled in different configurations using 3D printed connectors. The paper investigates the potential of utilizing the simple sensory feedback from the flex and pressure sensors to make additional meaningful inferences regarding the contact state and grasped object size. We study the effect of the grasped object size and contact type on the combined feedback from the embedded flex sensors of opposing fingers. Our results show that a simple linear relationship exists between the grasped object size and the final flex sensor reading at fixed input conditions, despite the variation in object weight and contact type. Additionally, by simply monitoring the time series response from the flex sensor, contact can be detected by comparing the response to the known free-bending response at the same input conditions. Furthermore, by utilizing the measured internal pressure supplied to the soft fingers, it is possible to distinguish between power and pinch grasps, as the contact type affects the rate of change in the flex sensor readings against the internal pressure.},
    keywords = {ARRAY(0x558b538acdd0)},
    url = {http://eprints.lincoln.ac.uk/34713/}
    }
  • C. Fox, Use and citation of paper “fox et al (2018), ?when should the chicken cross the road? game theory for autonomous vehicle – human interactions conference paper?” by the law commission to review and potentially change the law of the uk on autonomous vehicles. cited in their consultation report, “automated vehicles: a joint preliminary consultation paper” on p174, ref 651., 2019.
    [BibTeX] [Abstract] [Download PDF]

    Topic of this consultation: The Centre for Connected and Automated Vehicles (CCAV) has asked the Law Commission of England and Wales and the Scottish Law Commission to examine options for regulating automated road vehicles. It is a three-year project, running from March 2018 to March 2021. This preliminary consultation paper focuses on the safety of passenger vehicles. Driving automation refers to a broad range of vehicle technologies. Examples range from widely-used technologies that assist human drivers (such as cruise control) to vehicles that drive themselves with no human intervention. We concentrate on automated driving systems which do not need human drivers for at least part of the journey. This paper looks at are three key themes. First, we consider how safety can be assured before and after automated driving systems are deployed. Secondly, we explore criminal and civil liability. Finally, we examine the need to adapt road rules for artificial intelligence.

    @misc{lirolem34922,
    month = {January},
    journal = {Automated Vehicles: A joint preliminary consultation paper},
    title = {Use and citation of paper "Fox et al (2018), ?When should the chicken cross the road? Game theory for autonomous vehicle - human interactions conference paper?" by the Law Commission to review and potentially change the law of the UK on autonomous vehicles. Cited in their consultation report, "Automated Vehicles: A joint preliminary consultation paper" on p174, ref 651.},
    author = {Charles Fox},
    year = {2019},
    url = {http://eprints.lincoln.ac.uk/34922/},
    keywords = {ARRAY(0x558b538abde8)},
    abstract = {Topic of this consultation: The Centre for Connected and Automated Vehicles (CCAV) has
    asked the Law Commission of England and Wales and the Scottish Law Commission to
    examine options for regulating automated road vehicles. It is a three-year project, running from
    March 2018 to March 2021. This preliminary consultation paper focuses on the safety of
    passenger vehicles.
    Driving automation refers to a broad range of vehicle technologies. Examples range from
    widely-used technologies that assist human drivers (such as cruise control) to vehicles that
    drive themselves with no human intervention. We concentrate on automated driving systems
    which do not need human drivers for at least part of the journey.
    This paper looks at are three key themes. First, we consider how safety can be assured before
    and after automated driving systems are deployed. Secondly, we explore criminal and civil
    liability. Finally, we examine the need to adapt road rules for artificial intelligence.}
    }
  • Q. Fu, H. Wang, C. Hu, and S. Yue, “Towards computational models and applications of insect visual systems for motion perception: a review,” Artificial life, vol. 25, iss. 3, 2019.
    [BibTeX] [Abstract] [Download PDF]

    Motion perception is a critical capability determining a variety of aspects of insects’ life, including avoiding predators, foraging and so forth. A good number of motion detectors have been identified in the insects’ visual pathways. Computational modelling of these motion detectors has not only been providing effective solutions to artificial intelligence, but also benefiting the understanding of complicated biological visual systems. These biological mechanisms through millions of years of evolutionary development will have formed solid modules for constructing dynamic vision systems for future intelligent machines. This article reviews the computational motion perception models originating from biological research of insects’ visual systems in the literature. These motion perception models or neural networks comprise the looming sensitive neuronal models of lobula giant movement detectors (LGMDs) in locusts, the translation sensitive neural systems of direction selective neurons (DSNs) in fruit flies, bees and locusts, as well as the small target motion detectors (STMDs) in dragonflies and hover flies. We also review the applications of these models to robots and vehicles. Through these modelling studies, we summarise the methodologies that generate different direction and size selectivity in motion perception. At last, we discuss about multiple systems integration and hardware realisation of these bio-inspired motion perception models.

    @article{lirolem35584,
    publisher = {MIT Press},
    journal = {Artificial life},
    volume = {25},
    year = {2019},
    title = {Towards Computational Models and Applications of Insect Visual Systems for Motion Perception: A Review},
    author = {Qinbing Fu and Hongxin Wang and Cheng Hu and Shigang Yue},
    number = {3},
    keywords = {ARRAY(0x558b5356be10)},
    abstract = {Motion perception is a critical capability determining a variety of aspects of insects' life, including avoiding predators, foraging and so forth. A good number of motion detectors have been identified in the insects' visual pathways. Computational modelling of these motion detectors has not only been providing effective solutions to artificial intelligence, but also benefiting the understanding of complicated biological visual systems. These biological mechanisms through millions of years of evolutionary development will have formed solid modules for constructing dynamic vision systems for future intelligent machines. This article reviews the computational motion perception models originating from biological research of insects' visual systems in the literature. These motion perception models or neural networks comprise the looming sensitive neuronal models of lobula giant movement detectors (LGMDs) in locusts, the translation sensitive neural systems of direction selective neurons (DSNs) in fruit flies, bees and locusts, as well as the small target motion detectors (STMDs) in dragonflies and hover flies. We also review the applications of these models to robots and vehicles. Through these modelling studies, we summarise the methodologies that generate different direction and size selectivity in motion perception. At last, we discuss about multiple systems integration and hardware realisation of these bio-inspired motion perception models.},
    url = {http://eprints.lincoln.ac.uk/35584/}
    }
  • Q. Fu, N. Bellotto, H. Wang, C. F. Rind, H. Wang, and S. Yue, “A visual neural network for robust collision perception in vehicle driving scenarios,” in 15th international conference on artificial intelligence applications and innovations, 2019.
    [BibTeX] [Abstract] [Download PDF]

    This research addresses the challenging problem of visual collision detection in very complex and dynamic real physical scenes, specifically, the vehicle driving scenarios. This research takes inspiration from a large-field looming sensitive neuron, i.e., the lobula giant movement detector (LGMD) in the locust’s visual pathways, which represents high spike frequency to rapid approaching objects. Building upon our previous models, in this paper we propose a novel inhibition mechanism that is capable of adapting to different levels of background complexity. This adaptive mechanism works effectively to mediate the local inhibition strength and tune the temporal latency of local excitation reaching the LGMD neuron. As a result, the proposed model is effective to extract colliding cues from complex dynamic visual scenes. We tested the proposed method using a range of stimuli including simulated movements in grating backgrounds and shifting of a natural panoramic scene, as well as vehicle crash video sequences. The experimental results demonstrate the proposed method is feasible for fast collision perception in real-world situations with potential applications in future autonomous vehicles.

    @inproceedings{lirolem35586,
    year = {2019},
    title = {A Visual Neural Network for Robust Collision Perception in Vehicle Driving Scenarios},
    author = {Qinbing Fu and Nicola Bellotto and Huatian Wang and F. Claire Rind and Hongxin Wang and Shigang Yue},
    booktitle = {15th International Conference on Artificial Intelligence Applications and Innovations},
    month = {May},
    url = {http://eprints.lincoln.ac.uk/35586/},
    keywords = {ARRAY(0x558b53806590)},
    abstract = {This research addresses the challenging problem of visual collision detection in very complex and dynamic real physical scenes, specifically, the vehicle driving scenarios. This research takes inspiration from a large-field looming sensitive neuron, i.e., the lobula giant movement detector (LGMD) in the locust's visual pathways, which represents high spike frequency to rapid approaching objects. Building upon our previous models, in this paper we propose a novel inhibition mechanism that is capable of adapting to different levels of background complexity. This adaptive mechanism works effectively to mediate the local inhibition strength and tune the temporal latency of local excitation reaching the LGMD neuron. As a result, the proposed model is effective to extract colliding cues from complex dynamic visual scenes. We tested the proposed method using a range of stimuli including simulated movements in grating backgrounds and shifting of a natural panoramic scene, as well as vehicle crash video sequences. The experimental results demonstrate the proposed method is feasible for fast collision perception in real-world situations with potential applications in future autonomous vehicles.}
    }
  • A. Gabriel, S. Cosar, N. Bellotto, and P. Baxter, “A dataset for action recognition in the wild,” in Towards autonomous robotic systems, 2019, p. 362–374.
    [BibTeX] [Abstract] [Download PDF]

    The development of autonomous robots for agriculture depends on a successful approach to recognize user needs as well as datasets reflecting the characteristics of the domain. Available datasets for 3D Action Recognition generally feature controlled lighting and framing while recording subjects from the front. They mostly reflect good recording conditions and therefore fail to account for the highly variable conditions the robot would have to work with in the field, e.g. when providing in-field logistic support for human fruit pickers as in our scenario. Existing work on Intention Recognition mostly labels plans or actions as intentions, but neither of those fully capture the extend of human intent. In this work, we argue for a holistic view on human Intention Recognition and propose a set of recording conditions, gestures and behaviors that better reflect the environment and conditions an agricultural robot might find itself in. We demonstrate the utility of the dataset by means of evaluating two human detection methods: bounding boxes and skeleton extraction.

    @inproceedings{lirolem36395,
    year = {2019},
    author = {Alexander Gabriel and Serhan Cosar and Nicola Bellotto and Paul Baxter},
    title = {A Dataset for Action Recognition in the Wild},
    publisher = {Springer},
    volume = {11649},
    booktitle = {Towards Autonomous Robotic Systems},
    pages = {362--374},
    month = {June},
    url = {http://eprints.lincoln.ac.uk/36395/},
    abstract = {The development of autonomous robots for agriculture depends on a successful approach to recognize user needs as well as datasets reflecting the characteristics of the domain. Available datasets for 3D Action Recognition generally feature controlled lighting and framing while recording subjects from the front. They mostly reflect good recording conditions and therefore fail to account for the highly variable conditions the robot would have to work with in the field, e.g. when providing in-field logistic support for human fruit pickers as in our scenario. Existing work on Intention Recognition mostly labels plans or actions as intentions, but neither of those fully capture the extend of human intent. In this work, we argue for a holistic view on human Intention Recognition and propose a set of recording conditions, gestures and behaviors that better reflect the environment and conditions an agricultural robot might find itself in. We demonstrate the utility of the dataset by means of evaluating two human detection methods: bounding boxes and skeleton extraction.},
    keywords = {ARRAY(0x558b535a6190)}
    }
  • A. Gabriel, N. Bellotto, and P. Baxter, “Towards a dataset of activities for action recognition in open fields,” in 2nd uk-ras robotics and autonomous systems conference, 2019, p. 64–67.
    [BibTeX] [Abstract] [Download PDF]

    In an agricultural context, having autonomous robots that can work side-by-side with human workers provide a range of productivity benefits. In order for this to be achieved safely and effectively, these autonomous robots require the ability to understand a range of human behaviors in order to facilitate task communication and coordination. The recognition of human actions is a key part of this, and is the focus of this paper. Available datasets for Action Recognition generally feature controlled lighting and framing while recording subjects from the front. They mostly reflect good recording conditions but fail to model the data a robot will have to work with in the field, such as varying distance and lighting conditions. In this work, we propose a set of recording conditions, gestures and behaviors that better reflect the environment an agricultural robot might find itself in and record a dataset with a range of sensors that demonstrate these conditions.

    @inproceedings{lirolem36201,
    year = {2019},
    author = {Alexander Gabriel and Nicola Bellotto and Paul Baxter},
    title = {Towards a Dataset of Activities for Action Recognition in Open Fields},
    publisher = {UK-RAS},
    booktitle = {2nd UK-RAS Robotics and Autonomous Systems Conference},
    pages = {64--67},
    month = {January},
    url = {http://eprints.lincoln.ac.uk/36201/},
    abstract = {In an agricultural context, having autonomous robots that can work side-by-side with human workers provide a range of productivity benefits. In order for this to be achieved safely and effectively, these autonomous robots require the ability to understand a range of human behaviors in order to facilitate task communication and coordination. The recognition of human actions is a key part of this, and is the focus of this paper. Available datasets for Action Recognition generally feature controlled lighting and framing while recording subjects from the front. They mostly reflect good recording conditions but fail to model the data a robot will have to work with in the field, such as varying distance and lighting conditions. In this work, we propose a set of recording conditions, gestures and behaviors that better reflect the environment an agricultural
    robot might find itself in and record a dataset with a range of sensors that demonstrate these conditions.},
    keywords = {ARRAY(0x558b535e2f88)}
    }
  • B. Grieve, T. Duckett, M. Collison, L. Boyd, J. West, Y. Hujun, F. Arvin, and S. Pearson, “The challenges posed by global broadacre crops in delivering smart agri-robotic solutions: a fundamental rethink is required.,” Global food security, vol. 23, p. 116–124, 2019.
    [BibTeX] [Abstract] [Download PDF]

    Threats to global food security from multiple sources, such as population growth, ageing farming populations, meat consumption trends, climate-change effects on abiotic and biotic stresses, the environmental impacts of agriculture are well publicised. In addition, with ever increasing tolerance of pest, diseases and weeds there is growing pressure on traditional crop genetic and protective chemistry technologies of the ?Green Revolution?. To ease the burden of these challenges, there has been a move to automate and robotise aspects of the farming process. This drive has focussed typically on higher value sectors, such as horticulture and viticulture, that have relied on seasonal manual labour to maintain produce supply. In developed economies, and increasingly developing nations, pressure on labour supply has become unsustainable and forced the need for greater mechanisation and higher labour productivity. This paper creates the case that for broadacre crops, such as cereals, a wholly new approach is necessary, requiring the establishment of an integrated biology & physical engineering infrastructure, which can work in harmony with current breeding, chemistry and agronomic solutions. For broadacre crops the driving pressure is to sustainably intensify production; increase yields and/or productivity whilst reducing environmental impact. Additionally, our limited understanding of the complex interactions between the variations in pests, weeds, pathogens, soils, water, environment and crops is inhibiting growth in resource productivity and creating yield gaps. We argue that for agriculture to deliver knowledge based sustainable intensification requires a new generation of Smart Technologies, which combine sensors and robotics with localised and/or cloud-based Artificial Intelligence (AI).

    @article{lirolem35842,
    publisher = {Elsevier},
    volume = {23},
    title = {The challenges posed by global broadacre crops in delivering smart agri-robotic solutions: A fundamental rethink is required.},
    pages = {116--124},
    month = {December},
    journal = {Global Food Security},
    year = {2019},
    author = {Bruce Grieve and Tom Duckett and Martin Collison and Lesley Boyd and Jon West and Yin Hujun and Farshad Arvin and Simon Pearson},
    url = {http://eprints.lincoln.ac.uk/35842/},
    abstract = {Threats to global food security from multiple sources, such as population growth, ageing farming populations, meat consumption trends, climate-change effects on abiotic and biotic stresses, the environmental impacts of agriculture are well publicised. In addition, with ever increasing tolerance of pest, diseases and weeds there is growing pressure on traditional crop genetic and protective chemistry technologies of the ?Green Revolution?. To ease the burden of these challenges, there has been a move to automate and robotise aspects of the farming process. This drive has focussed typically on higher value sectors, such as horticulture and viticulture, that have relied on seasonal manual labour to maintain produce supply. In developed economies, and increasingly developing nations, pressure on labour supply has become unsustainable and forced the need for greater mechanisation and higher labour productivity. This paper creates the case that for broadacre crops, such as cereals, a wholly new approach is necessary, requiring the establishment of an integrated biology \& physical engineering infrastructure, which can work in harmony with current breeding, chemistry and agronomic solutions. For broadacre crops the driving pressure is to sustainably intensify production; increase yields and/or productivity whilst reducing environmental impact. Additionally, our limited understanding of the complex interactions between the variations in pests, weeds, pathogens, soils, water, environment and crops is inhibiting growth in resource productivity and creating yield gaps. We argue that for agriculture to deliver knowledge based sustainable intensification requires a new generation of Smart Technologies, which combine sensors and robotics with localised and/or cloud-based Artificial Intelligence (AI).},
    keywords = {ARRAY(0x558b532821c8)}
    }
  • M. Hüttenrauch, S. Adrian, and G. Neumann, “Deep reinforcement learning for swarm systems,” Journal of machine learning research, vol. 20, iss. 54, p. 1–31, 2019.
    [BibTeX] [Abstract] [Download PDF]

    Recently, deep reinforcement learning (RL) methods have been applied successfully to multi-agent scenarios. Typically, the observation vector for decentralized decision making is represented by a concatenation of the (local) information an agent gathers about other agents. However, concatenation scales poorly to swarm systems with a large number of homogeneous agents as it does not exploit the fundamental properties inherent to these systems: (i) the agents in the swarm are interchangeable and (ii) the exact number of agents in the swarm is irrelevant. Therefore, we propose a new state representation for deep multi-agent RL based on mean embeddings of distributions, where we treat the agents as samples and use the empirical mean embedding as input for a decentralized policy. We define different feature spaces of the mean embedding using histograms, radial basis functions and neural networks trained end-to-end. We evaluate the representation on two well-known problems from the swarm literature in a globally and locally observable setup. For the local setup we furthermore introduce simple communication protocols. Of all approaches, the mean embedding representation using neural network features enables the richest information exchange between neighboring agents, facilitating the development of complex collective strategies.

    @article{lirolem36281,
    title = {Deep Reinforcement Learning for Swarm Systems},
    volume = {20},
    publisher = {Journal of Machine Learning Research},
    month = {February},
    number = {54},
    pages = {1--31},
    author = {Maximilian H{\"u}ttenrauch and Sosic Adrian and Gerhard Neumann},
    year = {2019},
    journal = {Journal of Machine Learning Research},
    keywords = {ARRAY(0x558b539309e8)},
    abstract = {Recently, deep reinforcement learning (RL) methods have been applied successfully to multi-agent scenarios. Typically, the observation vector for decentralized decision making is represented by a concatenation of the (local) information an agent gathers about other agents. However, concatenation scales poorly to swarm systems with a large number of homogeneous agents as it does not exploit the fundamental properties inherent to these systems: (i) the agents in the swarm are interchangeable and (ii) the exact number of agents in the swarm is irrelevant. Therefore, we propose a new state representation for deep multi-agent RL based on mean embeddings of distributions, where we treat the agents as samples and use the empirical mean embedding as input for a decentralized policy. We define different feature spaces of the mean embedding using histograms, radial basis functions and neural networks trained end-to-end. We evaluate the representation on two well-known problems from the swarm literature in a globally and locally observable setup. For the local setup we furthermore introduce simple communication protocols. Of all approaches, the mean embedding representation using neural network features enables the richest information exchange between neighboring agents, facilitating the development of complex collective strategies.},
    url = {http://eprints.lincoln.ac.uk/36281/}
    }
  • S. Kottayil, P. Tsoleridis, K. Rossa, and C. Fox, “Investigation of driver route choice behaviour using bluetooth data,” in 15th world conference on transport research, 2019.
    [BibTeX] [Abstract] [Download PDF]

    Many local authorities use small-scale transport models to manage their transportation networks. These may assume drivers? behaviour to be rational in choosing the fastest route, and thus that all drivers behave the same given an origin and destination, leading to simplified aggregate flow models, fitted to anonymous traffic flow measurements. Recent price falls in traffic sensors, data storage, and compute power now enable Data Science to empirically test such assumptions, by using per-driver data to infer route selection from sensor observations and compare with optimal route selection. A methodology is presented using per-driver data to analyse driver route choice behaviour in transportation networks. Traffic flows on multiple measurable routes for origin destination pairs are compared based on the length of each route. A driver rationality index is defined by considering the shortest physical route between an origin-destination pair. The proposed method is intended to aid calibration of parameters used in traffic assignment models e.g. weights in generalized cost formulations or dispersion within stochastic user equilibrium models. The method is demonstrated using raw sensor datasets collected through Bluetooth sensors in the area of Chesterfield, Derbyshire, UK. The results for this region show that routes with a significant difference in lengths of their paths have the majority (71\%) of drivers using the optimal path but as the difference in length decreases, the probability of suboptimal route choice decreases (27\%). The methodology can be used for extended research considering the impact on route choice of other factors including travel time and road specific conditions.

    @inproceedings{lirolem34791,
    month = {May},
    author = {Sreedevi Kottayil and Panagiotis Tsoleridis and Kacper Rossa and Charles Fox},
    title = {Investigation of Driver Route Choice Behaviour using Bluetooth Data},
    year = {2019},
    booktitle = {15th World Conference on Transport Research},
    publisher = {Elsevier},
    url = {http://eprints.lincoln.ac.uk/34791/},
    abstract = {Many local authorities use small-scale transport models to manage their transportation networks. These may assume drivers? behaviour to be rational in choosing the fastest route, and thus that all drivers behave the same given an origin and destination, leading to simplified aggregate flow models, fitted to anonymous traffic flow measurements. Recent price falls in traffic sensors, data storage, and compute power now enable Data Science to empirically test such assumptions, by using per-driver data to infer route selection from sensor observations and compare with optimal route selection. A methodology is presented using per-driver data to analyse driver route choice behaviour in transportation networks. Traffic flows on multiple measurable routes for origin destination pairs are compared based on the length of each route. A driver rationality index is defined by considering the shortest physical route between an origin-destination pair. The proposed method is intended to aid calibration of parameters used in traffic assignment models e.g. weights in generalized cost formulations or dispersion within stochastic user equilibrium models. The method is demonstrated using raw sensor datasets collected through Bluetooth sensors in the area of Chesterfield, Derbyshire, UK. The results for this region show that routes with a significant difference in lengths of their paths have the majority (71\%) of drivers using the optimal path but as the difference in length decreases, the probability of suboptimal route choice decreases (27\%). The methodology can be used for extended research considering the impact on route choice of other factors including travel time and road specific conditions.},
    keywords = {ARRAY(0x558b535ddac0)}
    }
  • D. Laparidou, F. Curtis, K. Goher, A. Kucukyilmaz, M. Walker, J. Akanuwe, and N. Siriwardena, Patient, carer and staff perceptions of robotics in rehabilitation: protocol of a systematic review and qualitative meta-synthesisPROSPERO International prospective register of systematic reviews, 2019.
    [BibTeX] [Abstract] [Download PDF]

    Registration of a systematic review and qualitative meta-synthesis protocol.

    @misc{lirolem36226,
    month = {June},
    publisher = {PROSPERO International prospective register of systematic reviews},
    title = {Patient, carer and staff perceptions of robotics in rehabilitation: protocol of a systematic review and qualitative meta-synthesis},
    author = {Despina Laparidou and Ffion Curtis and Khaled Goher and Ayse Kucukyilmaz and Marion Walker and Joseph Akanuwe and Niro Siriwardena},
    year = {2019},
    keywords = {ARRAY(0x558b5378eef0)},
    abstract = {Registration of a systematic review and qualitative meta-synthesis protocol.},
    url = {http://eprints.lincoln.ac.uk/36226/}
    }
  • J. Lock, A. G. Tramontano, S. Ghidoni, and N. Bellotto, “Activis: mobile object detection and active guidance for people with visual impairments,” in Proc. of the int. conf. on image analysis and processing (iciap), 2019.
    [BibTeX] [Abstract] [Download PDF]

    The ActiVis project aims to deliver a mobile system that is able to guide a person with visual impairments towards a target object or area in an unknown indoor environment. For this, it uses new developments in object detection, mobile computing, action generation and human-computer interfacing to interpret the user’s surroundings and present effective guidance directions. Our approach to direction generation uses a Partially Observable Markov Decision Process (POMDP) to track the system’s state and output the optimal location to be investigated. This system includes an object detector and an audio-based guidance interface to provide a complete active search pipeline. The ActiVis system was evaluated in a set of experiments showing better performance than a simpler unguided case.

    @inproceedings{lirolem36413,
    booktitle = {Proc. of the Int. Conf. on Image Analysis and Processing (ICIAP)},
    year = {2019},
    author = {Jacobus Lock and A. G. Tramontano and S. Ghidoni and Nicola Bellotto},
    title = {ActiVis: Mobile Object Detection and Active Guidance for People with Visual Impairments},
    url = {http://eprints.lincoln.ac.uk/36413/},
    keywords = {ARRAY(0x558b5393a3d8)},
    abstract = {The ActiVis project aims to deliver a mobile system that is able to guide a person with visual impairments towards a target object or area in an unknown indoor environment. For this, it uses new developments in object detection, mobile computing, action generation and human-computer interfacing to interpret the user's surroundings and present effective guidance directions. Our approach to direction generation uses a Partially Observable Markov Decision Process (POMDP) to track the system's state and output the optimal location to be investigated. This system includes an object detector and an audio-based guidance interface to provide a complete active search pipeline. The ActiVis system was evaluated in a set of experiments showing better performance than a simpler unguided case.}
    }
  • J. Lock, G. Cielniak, and N. Bellotto, “Active object search with a mobile device for people with visual impairments,” in 14th international conference on computer vision theory and applications (visapp), 2019, p. 476–485.
    [BibTeX] [Abstract] [Download PDF]

    Modern smartphones can provide a multitude of services to assist people with visual impairments, and their cameras in particular can be useful for assisting with tasks, such as reading signs or searching for objects in unknown environments. Previous research has looked at ways to solve these problems by processing the camera’s video feed, but very little work has been done in actively guiding the user towards specific points of interest, maximising the effectiveness of the underlying visual algorithms. In this paper, we propose a control algorithm based on a Markov Decision Process that uses a smartphone?s camera to generate real-time instructions to guide a user towards a target object. The solution is part of a more general active vision application for people with visual impairments. An initial implementation of the system on a smartphone was experimentally evaluated with participants with healthy eyesight to determine the performance of the control algorithm. The results show the effectiveness of our solution and its potential application to help people with visual impairments find objects in unknown environments.

    @inproceedings{lirolem34596,
    publisher = {VISIGRAPP},
    booktitle = {14th International Conference on Computer Vision Theory and Applications (VISAPP)},
    year = {2019},
    title = {Active Object Search with a Mobile Device for People with Visual Impairments},
    author = {Jacobus Lock and Grzegorz Cielniak and Nicola Bellotto},
    pages = {476--485},
    url = {http://eprints.lincoln.ac.uk/34596/},
    keywords = {ARRAY(0x558b5356bdc8)},
    abstract = {Modern smartphones can provide a multitude of services to assist people with visual impairments, and their cameras in particular can be useful for assisting with tasks, such as reading signs or searching for objects in unknown environments. Previous research has looked at ways to solve these problems by processing the camera's video feed, but very little work has been done in actively guiding the user towards specific points of interest, maximising the effectiveness of the underlying visual algorithms. In this paper, we propose a control algorithm based on a Markov Decision Process that uses a smartphone?s camera to generate real-time instructions to guide a user towards a target object. The solution is part of a more general active vision application for people with visual impairments. An initial implementation of the system on a smartphone was experimentally evaluated with participants with healthy eyesight to determine the performance of the control algorithm. The results show the effectiveness of our solution and its potential application to help people with visual impairments find objects in unknown environments.}
    }
  • S. Molina, G. Cielniak, and T. Duckett, “Go with the flow: exploration and mapping of pedestrian flow patterns from partial observations,” in International conference on robotics and automation (icra), 2019.
    [BibTeX] [Abstract] [Download PDF]

    Understanding how people are likely to behave in an environment is a key requirement for efficient and safe robot navigation. However, mobile platforms are subject to spatial and temporal constraints, meaning that only partial observations of human activities are typically available to a robot, while the activity patterns of people in a given environment may also change at different times. To address these issues we present as the main contribution an exploration strategy for acquiring models of pedestrian flows, which decides not only the locations to explore but also the times when to explore them. The approach is driven by the uncertainty from multiple Poisson processes built from past observations. The approach is evaluated using two long-term pedestrian datasets, comparing its performance against uninformed exploration strategies. The results show that when using the uncertainty in the exploration policy, model accuracy increases, enabling faster learning of human motion patterns.

    @inproceedings{lirolem36396,
    title = {Go with the Flow: Exploration and Mapping of Pedestrian Flow Patterns from Partial Observations},
    author = {Sergi Molina and Grzegorz Cielniak and Tom Duckett},
    year = {2019},
    booktitle = {International Conference on Robotics and Automation (ICRA)},
    abstract = {Understanding how people are likely to behave in an environment is a key requirement for efficient and safe robot navigation. However, mobile platforms are subject to spatial and temporal constraints, meaning that only partial observations of human activities are typically available to a robot, while the activity patterns of people in a given environment may also change at different times. To address these issues we present as the main contribution an exploration strategy for acquiring models of pedestrian flows, which decides not only the locations to explore but also the times when to explore them. The approach is driven by the uncertainty from multiple Poisson processes built from past observations. The approach is evaluated using two long-term pedestrian datasets, comparing its performance against uninformed exploration strategies. The results show that when using the uncertainty in the exploration policy, model accuracy increases, enabling faster learning of human motion patterns.},
    keywords = {ARRAY(0x558b534096b8)},
    url = {http://eprints.lincoln.ac.uk/36396/}
    }
  • J. Pajarinen, H. L. Thai, R. Akrour, J. Peters, and G. Neumann, “Compatible natural gradient policy search,” Machine learning, 2019.
    [BibTeX] [Abstract] [Download PDF]

    Trust-region methods have yielded state-of-the-art results in policy search. A common approach is to use KL-divergence to bound the region of trust resulting in a natural gradient policy update. We show that the natural gradient and trust region optimization are equivalent if we use the natural parameterization of a standard exponential policy distribution in combination with compatible value function approximation. Moreover, we show that standard natural gradient updates may reduce the entropy of the policy according to a wrong schedule leading to premature convergence. To control entropy reduction we introduce a new policy search method called compatible policy search (COPOS) which bounds entropy loss. The experimental results show that COPOS yields state-of-the-art results in challenging continuous control tasks and in discrete partially observable tasks.

    @article{lirolem36283,
    title = {Compatible natural gradient policy search},
    author = {J. Pajarinen and H.L. Thai and R. Akrour and J. Peters and Gerhard Neumann},
    year = {2019},
    journal = {Machine Learning},
    publisher = {Springer},
    url = {http://eprints.lincoln.ac.uk/36283/},
    abstract = {Trust-region methods have yielded state-of-the-art results in policy search. A common approach is to use KL-divergence to bound the region of trust resulting in a natural gradient policy update. We show that the natural gradient and trust region optimization are equivalent if we use the natural parameterization of a standard exponential policy distribution in combination with compatible value function approximation. Moreover, we show that standard natural gradient updates may reduce the entropy of the policy according to a wrong schedule leading to premature convergence. To control entropy reduction we introduce a
    new policy search method called compatible policy search (COPOS) which bounds entropy loss. The experimental results show that COPOS yields state-of-the-art results in challenging continuous control tasks and in discrete partially observable tasks.},
    keywords = {ARRAY(0x558b5393a438)}
    }
  • L. Sun, C. Zhao, Z. Yan, P. Liu, T. Duckett, and R. Stolkin, “A novel weakly-supervised approach for rgb-d-based nuclear waste object detection and categorization,” Ieee sensors journal, vol. 19, iss. 9, p. 3487–3500, 2019.
    [BibTeX] [Abstract] [Download PDF]

    This paper addresses the problem of RGBD-based detection and categorization of waste objects for nuclear de- commissioning. To enable autonomous robotic manipulation for nuclear decommissioning, nuclear waste objects must be detected and categorized. However, as a novel industrial application, large amounts of annotated waste object data are currently unavailable. To overcome this problem, we propose a weakly-supervised learning approach which is able to learn a deep convolutional neural network (DCNN) from unlabelled RGBD videos while requiring very few annotations. The proposed method also has the potential to be applied to other household or industrial applications. We evaluate our approach on the Washington RGB- D object recognition benchmark, achieving the state-of-the-art performance among semi-supervised methods. More importantly, we introduce a novel dataset, i.e. Birmingham nuclear waste simulants dataset, and evaluate our proposed approach on this novel industrial object recognition challenge. We further propose a complete real-time pipeline for RGBD-based detection and categorization of nuclear waste simulants. Our weakly-supervised approach has demonstrated to be highly effective in solving a novel RGB-D object detection and recognition application with limited human annotations.

    @article{lirolem35699,
    author = {Li Sun and Cheng Zhao and Zhi Yan and Pengcheng Liu and Tom Duckett and Rustam Stolkin},
    year = {2019},
    journal = {IEEE Sensors Journal},
    title = {A Novel Weakly-supervised approach for RGB-D-based Nuclear Waste Object Detection and Categorization},
    volume = {19},
    publisher = {IEEE},
    month = {May},
    number = {9},
    pages = {3487--3500},
    abstract = {This paper addresses the problem of RGBD-based detection and categorization of waste objects for nuclear de- commissioning. To enable autonomous robotic manipulation for nuclear decommissioning, nuclear waste objects must be detected and categorized. However, as a novel industrial application, large amounts of annotated waste object data are currently unavailable. To overcome this problem, we propose a weakly-supervised learning approach which is able to learn a deep convolutional neural network (DCNN) from unlabelled RGBD videos while requiring very few annotations. The proposed method also has the potential to be
    applied to other household or industrial applications. We evaluate our approach on the Washington RGB- D object recognition benchmark, achieving the state-of-the-art performance among semi-supervised methods. More importantly, we introduce a novel dataset, i.e. Birmingham nuclear waste simulants dataset, and evaluate our proposed approach on this novel industrial object recognition challenge. We further propose a complete real-time pipeline for RGBD-based detection and categorization of nuclear waste simulants. Our weakly-supervised approach has demonstrated to be highly effective in solving a novel RGB-D object
    detection and recognition application with limited human annotations.},
    keywords = {ARRAY(0x558b5351a338)},
    url = {http://eprints.lincoln.ac.uk/35699/}
    }
  • X. Sun, T. liu, C. Hu, Q. Fu, and S. Yue, “Colcos\ensuremath\Phi: a multiple pheromone communication system for swarm robotics and social insects research,” in The 2019 ieee international conference on advanced robotics and mechatronics (icarm), 2019.
    [BibTeX] [Abstract] [Download PDF]

    In the last few decades we have witnessed how the pheromone of social insect has become a rich inspiration source of swarm robotics. By utilising the virtual pheromone in physical swarm robot system to coordinate individuals and realise direct/indirect inter-robot communications like the social insect, stigmergic behaviour has emerged. However, many studies only take one single pheromone into account in solving swarm problems, which is not the case in real insects. In the real social insect world, diverse behaviours, complex collective performances and ?exible transition from one state to another are guided by different kinds of pheromones and their interactions. Therefore, whether multiple pheromone based strategy can inspire swarm robotics research, and inversely how the performances of swarm robots controlled by multiple pheromones bring inspirations to explain the social insects? behaviours will become an interesting question. Thus, to provide a reliable system to undertake the multiple pheromone study, in this paper, we speci?cally proposed and realised a multiple pheromone communication system called ColCOS{\ensuremath{\Phi}}. This system consists of a virtual pheromone sub-system wherein the multiple pheromone is represented by a colour image displayed on a screen, and the micro-robots platform designed for swarm robotics applications. Two case studies are undertaken to verify the effectiveness of this system: one is the multiple pheromone based on an ant?s forage and another is the interactions of aggregation and alarm pheromones. The experimental results demonstrate the feasibility of ColCOS{\ensuremath{\Phi}} and its great potential in directing swarm robotics and social insects research.

    @inproceedings{lirolem36187,
    month = {July},
    year = {2019},
    title = {ColCOS{\ensuremath{\Phi}}: A Multiple Pheromone Communication System for Swarm Robotics and Social Insects Research},
    author = {Xuelong Sun and Tian liu and Cheng Hu and Qinbing Fu and Shigang Yue},
    publisher = {IEEE},
    booktitle = {The 2019 IEEE International Conference on Advanced Robotics and Mechatronics (ICARM)},
    url = {http://eprints.lincoln.ac.uk/36187/},
    keywords = {ARRAY(0x558b532da7a8)},
    abstract = {In the last few decades we have witnessed how the pheromone of social insect has become a rich inspiration source of swarm robotics. By utilising the virtual pheromone in physical swarm robot system to coordinate individuals and realise direct/indirect inter-robot communications like the social insect, stigmergic behaviour has emerged. However, many studies only take one single pheromone into account in solving swarm problems, which is not the case in real insects. In the real social insect world, diverse behaviours, complex collective performances and ?exible transition from one state to another are guided by different kinds of pheromones and their interactions. Therefore, whether multiple pheromone based strategy can inspire swarm robotics research, and inversely how the performances of swarm robots controlled by multiple pheromones bring inspirations to explain the social insects? behaviours will become an interesting question. Thus, to provide a reliable system to undertake the multiple pheromone study, in this paper, we speci?cally proposed and realised a multiple pheromone communication system called ColCOS{\ensuremath{\Phi}}. This system consists of a virtual pheromone sub-system wherein the multiple pheromone is represented by a colour image displayed on a screen, and the micro-robots platform designed for swarm robotics applications. Two case studies are undertaken to verify the effectiveness of this system: one is the multiple pheromone based on an ant?s forage and another is the interactions of aggregation and alarm pheromones. The experimental results demonstrate the feasibility of ColCOS{\ensuremath{\Phi}} and its great potential in directing swarm robotics and social insects research.}
    }
  • H. Wang, J. Peng, X. Zheng, and S. Yue, “A robust visual system for small target motion detection against cluttered moving backgrounds,” Ieee transactions on neural networks and learning systems, 2019.
    [BibTeX] [Abstract] [Download PDF]

    Monitoring small objects against cluttered moving backgrounds is a huge challenge to future robotic vision systems. As a source of inspiration, insects are quite apt at searching for mates and tracking prey, which always appear as small dim speckles in the visual field. The exquisite sensitivity of insects for small target motion, as revealed recently, is coming from a class of specific neurons called small target motion detectors (STMDs). Although a few STMD-based models have been proposed, these existing models only use motion information for small target detection and cannot discriminate small targets from small-target-like background features (named fake features). To address this problem, this paper proposes a novel visual system model (STMD+) for small target motion detection, which is composed of four subsystems–ommatidia, motion pathway, contrast pathway, and mushroom body. Compared with the existing STMD-based models, the additional contrast pathway extracts directional contrast from luminance signals to eliminate false positive background motion. The directional contrast and the extracted motion information by the motion pathway are integrated into the mushroom body for small target discrimination. Extensive experiments showed the significant and consistent improvements of the proposed visual system model over the existing STMD-based models against fake features.

    @article{lirolem36114,
    month = {May},
    journal = {IEEE Transactions on Neural Networks and Learning Systems},
    publisher = {IEEE},
    year = {2019},
    title = {A Robust Visual System for Small Target Motion Detection Against Cluttered Moving Backgrounds},
    author = {Hongxin Wang and Jigen Peng and Xuqiang Zheng and Shigang Yue},
    keywords = {ARRAY(0x558b5360aba0)},
    abstract = {Monitoring small objects against cluttered moving backgrounds is a huge challenge to future robotic vision systems. As a source of inspiration, insects are quite apt at searching for mates and tracking prey, which always appear as small dim speckles in the visual field. The exquisite sensitivity of insects for small target motion, as revealed recently, is coming from a class of specific neurons called small target motion detectors (STMDs). Although a few STMD-based models have been proposed, these existing models only use motion information for small target detection and cannot discriminate small targets from small-target-like background features (named fake features). To address this problem, this paper proposes a novel visual system model (STMD+) for small target motion detection, which is composed of four subsystems--ommatidia, motion pathway, contrast pathway, and mushroom body. Compared with the existing STMD-based models, the additional contrast pathway extracts directional contrast from luminance signals to eliminate false positive background motion. The directional contrast and the extracted motion information by the motion pathway are integrated into the mushroom body for small target discrimination. Extensive experiments showed the significant and consistent improvements of the proposed visual system model over the existing STMD-based models against fake features.},
    url = {http://eprints.lincoln.ac.uk/36114/}
    }
  • H. Wang, J. Peng, Q. Fu, H. Wang, and S. Yue, “Visual cue integration for small target motion detection in natural cluttered backgrounds,” in The 2019 international joint conference on neural networks (ijcnn), 2019.
    [BibTeX] [Abstract] [Download PDF]

    The robust detection of small targets against cluttered background is important for future arti?cial visual systems in searching and tracking applications. The insects? visual systems have demonstrated excellent ability to avoid predators, ?nd prey or identify conspeci?cs ? which always appear as small dim speckles in the visual ?eld. Build a computational model of the insects? visual pathways could provide effective solutions to detect small moving targets. Although a few visual system models have been proposed, they only make use of small-?eld visual features for motion detection and their detection results often contain a number of false positives. To address this issue, we develop a new visual system model for small target motion detection against cluttered moving backgrounds. Compared to the existing models, the small-?eld and wide-?eld visual features are separately extracted by two motion-sensitive neurons to detect small target motion and background motion. These two types of motion information are further integrated to ?lter out false positives. Extensive experiments showed that the proposed model can outperform the existing models in terms of detection rates.

    @inproceedings{lirolem35684,
    month = {July},
    title = {Visual Cue Integration for Small Target Motion Detection in Natural Cluttered Backgrounds},
    author = {Hongxin Wang and Jigen Peng and Qinbing Fu and Huatian Wang and Shigang Yue},
    year = {2019},
    booktitle = {The 2019 International Joint Conference on Neural Networks (IJCNN)},
    publisher = {IEEE},
    abstract = {The robust detection of small targets against cluttered background is important for future arti?cial visual systems in searching and tracking applications. The insects? visual systems have demonstrated excellent ability to avoid predators, ?nd prey or identify conspeci?cs ? which always appear as small dim speckles in the visual ?eld. Build a computational model of the insects? visual pathways could provide effective solutions to detect small moving targets. Although a few visual system models have been proposed, they only make use of small-?eld visual features for motion detection and their detection results often contain a number of false positives. To address this issue, we develop a new visual system model for small target motion detection against cluttered moving backgrounds. Compared to the existing models, the small-?eld and wide-?eld visual features are separately extracted by two motion-sensitive neurons to detect small target motion and background motion. These two types of motion information are further integrated to ?lter out false positives. Extensive experiments showed that the proposed model can outperform the existing models in terms of detection rates.},
    keywords = {ARRAY(0x558b538f42a0)},
    url = {http://eprints.lincoln.ac.uk/35684/}
    }
  • H. Wang, Q. Fu, H. Wang, J. Peng, P. Baxter, C. Hu, and S. Yue, “Angular velocity estimation of image motion mimicking the honeybee tunnel centring behaviour,” in The 2019 international joint conference on neural networks, 2019.
    [BibTeX] [Abstract] [Download PDF]

    Insects use visual information to estimate angular velocity of retinal image motion, which determines a variety of ?ight behaviours including speed regulation, tunnel centring and visual navigation. For angular velocity estimation, honeybees show large spatial-independence against visual stimuli, whereas the previous models have not ful?lled such an ability. To address this issue, we propose a bio-plausible model for estimating the image motion velocity based on behavioural experiments of the honeybee ?ying through patterned tunnels. The proposed model contains mainly three parts, the texture estimation layer for spatial information extraction, the delay-and-correlate layer for temporal information extraction and the decoding layer for angular velocity estimation. This model produces responses that are largely independent of the spatial frequency in grating experiments. And the model has been implemented in a virtual bee for tunnel centring simulations. The results coincide with both electro-physiological neuron spike and behavioural path recordings, which indicates our proposed method provides a better explanation of the honeybee?s image motion detection mechanism guiding the tunnel centring behaviour.

    @inproceedings{lirolem35685,
    booktitle = {The 2019 International Joint Conference on Neural Networks},
    publisher = {IEEE},
    title = {Angular Velocity Estimation of Image Motion Mimicking the Honeybee Tunnel Centring Behaviour},
    author = {Huatian Wang and Qinbing Fu and Hongxin Wang and Jigen Peng and Paul Baxter and Cheng Hu and Shigang Yue},
    year = {2019},
    month = {July},
    abstract = {Insects use visual information to estimate angular velocity of retinal image motion, which determines a variety of ?ight behaviours including speed regulation, tunnel centring and visual navigation. For angular velocity estimation, honeybees show large spatial-independence against visual stimuli, whereas the previous models have not ful?lled such an ability. To address this issue, we propose a bio-plausible model for estimating the image motion velocity based on behavioural experiments of the honeybee ?ying through patterned tunnels. The proposed model contains mainly three parts, the texture estimation layer for spatial information extraction, the delay-and-correlate layer for temporal information extraction and the decoding layer for angular velocity estimation. This model produces responses that are largely independent of the spatial frequency in grating experiments. And the model has been implemented in a virtual bee for tunnel centring simulations. The results coincide with both electro-physiological neuron spike and behavioural path recordings, which indicates our proposed method provides a better explanation of the honeybee?s image motion detection mechanism guiding the tunnel centring behaviour.},
    keywords = {ARRAY(0x558b536af350)},
    url = {http://eprints.lincoln.ac.uk/35685/}
    }
  • H. Wang, Q. Fu, H. Wang, J. Peng, and S. Yue, “Constant angular velocity regulation for visually guided terrain following,” in 15th international conference on artificial intelligence applications and innovations, 2019.
    [BibTeX] [Abstract] [Download PDF]

    Insects use visual cues to control their flight behaviours. By estimating the angular velocity of the visual stimuli and regulating it to a constant value, honeybees can perform a terrain following task which keeps the certain height above the undulated ground. For mimicking this behaviour in a bio-plausible computation structure, this paper presents a new angular velocity decoding model based on the honeybee’s behavioural experiments. The model consists of three parts, the texture estimation layer for spatial information extraction, the motion detection layer for temporal information extraction and the decoding layer combining information from pervious layers to estimate the angular velocity. Compared to previous methods on this field, the proposed model produces responses largely independent of the spatial frequency and contrast in grating experiments. The angular velocity based control scheme is proposed to implement the model into a bee simulated by the game engine Unity. The perfect terrain following above patterned ground and successfully flying over irregular textured terrain show its potential for micro unmanned aerial vehicles’ terrain following.

    @inproceedings{lirolem35595,
    author = {Huatian Wang and Qinbing Fu and Hongxin Wang and Jigen Peng and Shigang Yue},
    title = {Constant Angular Velocity Regulation for Visually Guided Terrain Following},
    year = {2019},
    booktitle = {15th International Conference on Artificial Intelligence Applications and Innovations},
    url = {http://eprints.lincoln.ac.uk/35595/},
    keywords = {ARRAY(0x558b537fa328)},
    abstract = {Insects use visual cues to control their flight behaviours. By estimating the angular velocity of the visual stimuli and regulating it to a constant value, honeybees can perform a terrain following task which keeps the certain height above the undulated ground. For mimicking this behaviour in a bio-plausible computation structure, this paper presents a new angular velocity decoding model based on the honeybee's behavioural experiments. The model consists of three parts, the texture estimation layer for spatial information extraction, the motion detection layer for temporal information extraction and the decoding layer combining information from pervious layers to estimate the angular velocity. Compared to previous methods on this field, the proposed model produces responses largely independent of the spatial frequency and contrast in grating experiments. The angular velocity based control scheme is proposed to implement the model into a bee simulated by the game engine Unity. The perfect terrain following above patterned ground and successfully flying over irregular textured terrain show its potential for micro unmanned aerial vehicles' terrain following.}
    }
  • C. Zhao, L. Sun, P. Purkait, T. Duckett, and R. Stolkin, “Learning monocular visual odometry with dense 3d mapping from dense 3d flow,” in 2018 ieee/rsj international conference on intelligent robots and systems (iros), 2019.
    [BibTeX] [Abstract] [Download PDF]

    This paper introduces a fully deep learning approach to monocular SLAM, which can perform simultaneous localization using a neural network for learning visual odometry (L-VO) and dense 3D mapping. Dense 2D flow and a depth image are generated from monocular images by sub-networks, which are then used by a 3D flow associated layer in the L-VO network to generate dense 3D flow. Given this 3D flow, the dual-stream L-VO network can then predict the 6DOF relative pose and furthermore reconstruct the vehicle trajectory. In order to learn the correlation between motion directions, the Bivariate Gaussian modeling is employed in the loss function. The L-VO network achieves an overall performance of 2.68 \% for average translational error and 0.0143?/m for average rotational error on the KITTI odometry benchmark. Moreover, the learned depth is leveraged to generate a dense 3D map. As a result, an entire visual SLAM system, that is, learning monocular odometry combined with dense 3D mapping, is achieved.

    @inproceedings{lirolem36001,
    month = {January},
    booktitle = {2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
    publisher = {IEEE},
    title = {Learning Monocular Visual Odometry with Dense 3D Mapping from Dense 3D Flow},
    author = {Cheng Zhao and Li Sun and Pulak Purkait and Tom Duckett and Rustam Stolkin},
    year = {2019},
    url = {http://eprints.lincoln.ac.uk/36001/},
    keywords = {ARRAY(0x558b538c08f8)},
    abstract = {This paper introduces a fully deep learning approach to monocular SLAM, which can perform simultaneous localization using a neural network for learning visual odometry (L-VO) and dense 3D mapping. Dense 2D flow and a depth image are generated from monocular images by sub-networks, which are then used by a 3D flow associated layer in the L-VO network to generate dense 3D flow. Given this 3D flow, the dual-stream L-VO network can then predict the 6DOF relative pose and furthermore reconstruct the vehicle trajectory. In order to learn the correlation between motion directions, the Bivariate Gaussian modeling is employed in the loss function. The L-VO network achieves an overall performance of 2.68 \% for average translational error and 0.0143?/m for average rotational error on the KITTI odometry benchmark. Moreover, the learned depth is leveraged to generate a dense 3D map. As a result, an entire visual SLAM system, that is, learning monocular odometry combined with dense 3D mapping, is achieved.}
    }
  • J. Zhao, X. Ma, Q. Fu, C. Hu, and S. Yue, “An lgmd based competitive collision avoidance strategy for uav,” in The 15th international conference on artificial intelligence applications and innovations, 2019.
    [BibTeX] [Abstract] [Download PDF]

    Building a reliable and e?cient collision avoidance system for unmanned aerial vehicles (UAVs) is still a challenging problem. This research takes inspiration from locusts, which can ?y in dense swarms for hundreds of miles without collision. In the locust?s brain, a visual pathway of LGMD-DCMD (lobula giant movement detector and descending contra-lateral motion detector) has been identi?ed as collision perception system guiding fast collision avoidance for locusts, which is ideal for designing arti?cial vision systems. However, there is very few works investigating its potential in real-world UAV applications. In this paper, we present an LGMD based competitive collision avoidance method for UAV indoor navigation. Compared to previous works, we divided the UAV?s ?eld of view into four sub?elds each handled by an LGMD neuron. Therefore, four individual competitive LGMDs (C-LGMD) compete for guiding the directional collision avoidance of UAV. With more degrees of freedom compared to ground robots and vehicles, the UAV can escape from collision along four cardinal directions (e.g. the object approaching from the left-side triggers a rightward shifting of the UAV). Our proposed method has been validated by both simulations and real-time quadcopter arena experiments.

    @inproceedings{lirolem35691,
    month = {May},
    publisher = {Springer},
    booktitle = {The 15th International Conference on Artificial Intelligence Applications and Innovations},
    year = {2019},
    author = {Jiannan Zhao and Xingzao Ma and Qinbing Fu and Cheng Hu and Shigang Yue},
    title = {An LGMD Based Competitive Collision Avoidance Strategy for UAV},
    url = {http://eprints.lincoln.ac.uk/35691/},
    keywords = {ARRAY(0x558b53484680)},
    abstract = {Building a reliable and e?cient collision avoidance system for unmanned aerial vehicles (UAVs) is still a challenging problem. This research takes inspiration from locusts, which can ?y in dense swarms for hundreds of miles without collision. In the locust?s brain, a visual pathway of LGMD-DCMD (lobula giant movement detector and descending contra-lateral motion detector) has been identi?ed as collision perception system guiding fast collision avoidance for locusts, which is ideal for designing arti?cial vision systems. However, there is very few works investigating its potential in real-world UAV applications. In this paper, we present an LGMD based competitive collision avoidance method for UAV indoor navigation. Compared to previous works, we divided the UAV?s ?eld of view into four sub?elds each handled by an LGMD neuron. Therefore, four individual competitive LGMDs (C-LGMD) compete for guiding the directional collision avoidance of UAV. With more degrees of freedom compared to ground robots and vehicles, the UAV can escape from collision along four cardinal directions (e.g. the object approaching from the left-side triggers a rightward shifting of the UAV). Our proposed method has been validated by both simulations and real-time quadcopter arena experiments.}
    }

2018

  • P. Bosilj, T. Duckett, and G. Cielniak, “Analysis of morphology-based features for classification of crop and weeds in precision agriculture,” Ieee robotics and automation letters, vol. 3, iss. 4, p. 2950–2956, 2018.
    [BibTeX] [Abstract] [Download PDF]

    Determining the types of vegetation present in an image is a core step in many precision agriculture tasks. In this paper, we focus on pixel-based approaches for classification of crops versus weeds, especially for complex cases involving overlapping plants and partial occlusion. We examine the benefits of multi-scale and content-driven morphology-based descriptors called Attribute Profiles. These are compared to state-of-the art keypoint descriptors with a fixed neighbourhood previously used in precision agriculture, namely Histograms of Oriented Gradients and Local Binary Patterns. The proposed classification technique is especially advantageous when coupled with morphology-based segmentation on a max-tree structure, as the same representation can be re-used for feature extraction. The robustness of the approach is demonstrated by an experimental evaluation on two datasets with different crop types. The proposed approach compared favourably to state-of-the-art approaches without an increase in computational complexity, while being able to provide descriptors at a higher resolution.

    @article{lirolem32371,
    title = {Analysis of morphology-based features for classification of crop and weeds in precision agriculture},
    volume = {3},
    publisher = {IEEE},
    month = {October},
    pages = {2950--2956},
    number = {4},
    author = {Petra Bosilj and Tom Duckett and Grzegorz Cielniak},
    year = {2018},
    journal = {IEEE Robotics and Automation Letters},
    url = {http://eprints.lincoln.ac.uk/32371/},
    abstract = {Determining the types of vegetation present in an image is a core step in many precision agriculture tasks. In this paper, we focus on pixel-based approaches for classification of crops versus weeds, especially for complex cases involving overlapping plants and partial occlusion. We examine the benefits of multi-scale and content-driven morphology-based descriptors called Attribute Profiles. These are compared to state-of-the art keypoint descriptors with a fixed neighbourhood previously used in precision agriculture, namely Histograms of Oriented Gradients and Local Binary Patterns. The proposed classification technique is especially advantageous when coupled with morphology-based segmentation on a max-tree structure, as the same representation can be re-used for feature extraction. The robustness of the approach is demonstrated by an experimental evaluation on two datasets with different crop types. The proposed approach compared favourably to state-of-the-art approaches without an increase in computational complexity, while being able to provide descriptors at a higher resolution.},
    keywords = {ARRAY(0x558b53756878)}
    }
  • F. Camara, O. Giles, M. Rothmuller, P. Rasmussen, A. Vendelbo-Larsen, G. Markkula, Y-M. Lee, N. Merat, and C. Fox, “Predicting pedestrian road-crossing assertiveness for autonomous vehicle control,” in 21st ieee international conference on intelligent transportation systems, 2018.
    [BibTeX] [Abstract] [Download PDF]

    Autonomous vehicles (AVs) must interact with other road users including pedestrians. Unlike passive environments, pedestrians are active agents having their own utilities and decisions, which must be inferred and predicted by AVs in order to control interactions with them and navigation around them. In particular, when a pedestrian wishes to cross the road in front of the vehicle at an unmarked crossing, the pedestrian and AV must compete for the space, which may be considered as a game-theoretic interaction in which one agent must yield to the other. To inform AV controllers in this setting, this study collects and analyses data from real-world human road crossings to determine what features of crossing behaviours are predictive about the level of assertiveness of pedestrians and of the eventual winner of the interactions. It presents the largest and most detailed data set of its kind known to us, and new methods to analyze and predict pedestrian-vehicle interactions based upon it. Pedestrian-vehicle interactions are decomposed into sequences of independent discrete events. We use probabilistic methods ? regression and decision tree regression ? and sequence analysis to analyze sets and sub-sequences of actions used by both pedestrians and human drivers while crossing at an intersection, to find common patterns of behaviour and to predict the winner of each interaction. We report on the particular features found to be predictive and which can thus be integrated into game- theoretic AV controllers to inform real-time interactions.

    @inproceedings{lirolem33089,
    title = {Predicting pedestrian road-crossing assertiveness for autonomous vehicle control},
    author = {F Camara and O Giles and M Rothmuller and PH Rasmussen and A Vendelbo-Larsen and G Markkula and Y-M Lee and N Merat and Charles Fox},
    year = {2018},
    booktitle = {21st IEEE International Conference on Intelligent Transportation Systems},
    publisher = {IEEE},
    month = {November},
    keywords = {ARRAY(0x558b534c4ac0)},
    abstract = {Autonomous vehicles (AVs) must interact with other
    road users including pedestrians. Unlike passive environments,
    pedestrians are active agents having their own utilities and
    decisions, which must be inferred and predicted by AVs in order
    to control interactions with them and navigation around them.
    In particular, when a pedestrian wishes to cross the road in
    front of the vehicle at an unmarked crossing, the pedestrian
    and AV must compete for the space, which may be considered
    as a game-theoretic interaction in which one agent must yield
    to the other. To inform AV controllers in this setting, this study
    collects and analyses data from real-world human road crossings
    to determine what features of crossing behaviours are predictive
    about the level of assertiveness of pedestrians and of the eventual
    winner of the interactions. It presents the largest and most
    detailed data set of its kind known to us, and new methods to
    analyze and predict pedestrian-vehicle interactions based upon
    it. Pedestrian-vehicle interactions are decomposed into sequences
    of independent discrete events. We use probabilistic methods ?
    regression and decision tree regression ? and sequence analysis
    to analyze sets and sub-sequences of actions used by both
    pedestrians and human drivers while crossing at an intersection,
    to find common patterns of behaviour and to predict the winner
    of each interaction. We report on the particular features found
    to be predictive and which can thus be integrated into game-
    theoretic AV controllers to inform real-time interactions.},
    url = {http://eprints.lincoln.ac.uk/33089/}
    }
  • H. Cuayahuitl, S. Ryu, D. Lee, and J. Kim, “A study on dialogue reward prediction for open-ended conversational agents,” in Neurips workshop on conversational ai, 2018.
    [BibTeX] [Abstract] [Download PDF]

    The amount of dialogue history to include in a conversational agent is often underestimated and/or set in an empirical and thus possibly naive way. This suggests that principled investigations into optimal context windows are urgently needed given that the amount of dialogue history and corresponding representations can play an important role in the overall performance of a conversational system. This paper studies the amount of history required by conversational agents for reliably predicting dialogue rewards. The task of dialogue reward prediction is chosen for investigating the effects of varying amounts of dialogue history and their impact on system performance. Experimental results using a dataset of 18K human-human dialogues report that lengthy dialogue histories of at least 10 sentences are preferred (25 sentences being the best in our experiments) over short ones, and that lengthy histories are useful for training dialogue reward predictors with strong positive correlations between target dialogue rewards and predicted ones.

    @inproceedings{lirolem34433,
    month = {December},
    author = {Heriberto Cuayahuitl and Seonghan Ryu and Donghyeon Lee and Jihie Kim},
    title = {A Study on Dialogue Reward Prediction for Open-Ended Conversational Agents},
    year = {2018},
    booktitle = {NeurIPS Workshop on Conversational AI},
    publisher = {arXiv},
    url = {http://eprints.lincoln.ac.uk/34433/},
    abstract = {The amount of dialogue history to include in a conversational agent is often underestimated and/or set in an empirical and thus possibly naive way. This suggests that principled investigations into optimal context windows are urgently needed given that the amount of dialogue history and corresponding representations can play an important role in the overall performance of a conversational system. This paper studies the amount of history required by conversational agents for reliably predicting dialogue rewards. The task of dialogue reward prediction is chosen for investigating the effects of varying amounts of dialogue history and their impact on system performance. Experimental results using a dataset of 18K human-human dialogues report that lengthy dialogue histories of at least 10 sentences are preferred (25 sentences being the best in our experiments) over short ones, and that lengthy histories are useful for training dialogue reward predictors with strong positive correlations between target dialogue rewards and predicted ones.},
    keywords = {ARRAY(0x558b537fbd80)}
    }
  • Q. Fu, C. Hu, J. Peng, and S. Yue, “Shaping the collision selectivity in a looming sensitive neuron model with parallel on and off pathways and spike frequency adaptation,” Neural networks, vol. 106, p. 127–143, 2018.
    [BibTeX] [Abstract] [Download PDF]

    Shaping the collision selectivity in vision-based artificial collision-detecting systems is still an open challenge. This paper presents a novel neuron model of a locust looming detector, i.e. the lobula giant movement detector (LGMD1), in order to provide effective solutions to enhance the collision selectivity of looming objects over other visual challenges. We propose an approach to model the biologically plausible mechanisms of ON and OFF pathways and a biophysical mechanism of spike frequency adaptation (SFA) in the proposed LGMD1 visual neural network. The ON and OFF pathways can separate both dark and light looming features for parallel spatiotemporal computations. This works effectively on perceiving a potential collision from dark or light objects that approach; such a bio-plausible structure can also separate LGMD1’s collision selectivity to its neighbouring looming detector – the LGMD2.The SFA mechanism can enhance the LGMD1’s collision selectivity to approaching objects rather than receding and translating stimuli, which is a significant improvement compared with similar LGMD1 neuron models. The proposed framework has been tested using off-line tests of synthetic and real-world stimuli, as well as on-line bio-robotic tests. The enhanced collision selectivity of the proposed model has been validated in systematic experiments. The computational simplicity and robustness of this work have also been verified by the bio-robotic tests, which demonstrates potential in building neuromorphic sensors for collision detection in both a fast and reliable manner.

    @article{lirolem31536,
    month = {December},
    pages = {127--143},
    title = {Shaping the collision selectivity in a looming sensitive neuron model with parallel ON and OFF pathways and spike frequency adaptation},
    volume = {106},
    publisher = {Elsevier for European Neural Network Society (ENNS)},
    author = {Qinbing Fu and Cheng Hu and Jigen Peng and Shigang Yue},
    year = {2018},
    journal = {Neural Networks},
    url = {http://eprints.lincoln.ac.uk/31536/},
    abstract = {Shaping the collision selectivity in vision-based artificial collision-detecting systems is still an open challenge. This paper presents a novel neuron model of a locust looming detector, i.e. the lobula giant movement detector (LGMD1), in order to provide effective solutions to enhance the collision selectivity of looming objects over other visual challenges. We propose an approach to model the biologically plausible mechanisms of ON and OFF pathways and a biophysical mechanism of spike frequency adaptation (SFA) in the proposed LGMD1 visual neural network. The ON and OFF pathways can separate both dark and light looming features for parallel spatiotemporal computations. This works effectively on perceiving a potential collision from dark or light objects that approach; such a bio-plausible structure can also separate LGMD1's collision selectivity to its neighbouring looming detector -- the LGMD2.The SFA mechanism can enhance the LGMD1's collision selectivity to approaching objects rather than receding and translating stimuli, which is a significant improvement compared with similar LGMD1 neuron models. The proposed framework has been tested using off-line tests of synthetic and real-world stimuli, as well as on-line bio-robotic tests. The enhanced collision selectivity of the proposed model has been validated in systematic experiments. The computational simplicity and robustness of this work have also been verified by the bio-robotic tests, which demonstrates potential in building neuromorphic sensors for collision detection in both a fast and reliable manner.},
    keywords = {ARRAY(0x558b537fbd98)}
    }
  • Q. Fu, N. Bellotto, C. Hu, and S. Yue, “Performance of a visual fixation model in an autonomous micro robot inspired by drosophila physiology,” in Ieee international conference on robotics and biomimetics, 2018.
    [BibTeX] [Abstract] [Download PDF]

    In nature, lightweight and low-powered insects are ideal model systems to study motion perception strategies. Understanding the underlying characteristics and functionality of insects? visual systems is not only attractive to neural system modellers, but also critical in providing effective solutions to future robotics. This paper presents a novel modelling of dynamic vision system inspired by Drosophila physiology for mimicking fast motion tracking and a closed-loop behavioural response to ?xation. The proposed model was realised on embedded system in an autonomous micro robot which has limited computational resources. A monocular camera was applied as the only motion sensing modality. Systematic experiments including open-loop and closed-loop bio-robotic tests validated the proposed visual ?xation model: the robot showed motion tracking and ?xation behaviours similarly to insects; the image processing frequency can maintain 25 {$\sim$} 45Hz. Arena tests also demonstrated a successful following behaviour aroused by ?xation in navigation.

    @inproceedings{lirolem33846,
    month = {December},
    booktitle = {IEEE INTERNATIONAL CONFERENCE ON ROBOTICS AND BIOMIMETICS},
    year = {2018},
    author = {Qinbing Fu and Nicola Bellotto and Cheng Hu and Shigang Yue},
    title = {Performance of a Visual Fixation Model in an Autonomous Micro Robot Inspired by Drosophila Physiology},
    url = {http://eprints.lincoln.ac.uk/33846/},
    abstract = {In nature, lightweight and low-powered insects are ideal model systems to study motion perception strategies. Understanding the underlying characteristics and functionality of insects? visual systems is not only attractive to neural system modellers, but also critical in providing effective solutions to future robotics. This paper presents a novel modelling of dynamic vision system inspired by Drosophila physiology for mimicking fast motion tracking and a closed-loop behavioural response to ?xation. The proposed model was realised on embedded system in an autonomous micro robot which has limited computational resources. A monocular camera was applied as the only motion sensing modality. Systematic experiments including open-loop and closed-loop bio-robotic tests validated the proposed visual ?xation model: the robot showed motion tracking and ?xation behaviours similarly to insects; the image processing frequency can maintain 25 {$\sim$} 45Hz. Arena tests also demonstrated a successful following behaviour aroused by ?xation in navigation.},
    keywords = {ARRAY(0x558b53941358)}
    }
  • S. Indurthi, S. Yu, S. Back, and H. Cuayahuitl, “Cut to the chase: a context zoom-in network for reading comprehension,” in Empirical methods in natural language processing (emnlp), 2018.
    [BibTeX] [Abstract] [Download PDF]

    In recent years many deep neural networks have been proposed to solve Reading Comprehension (RC) tasks. Most of these models suffer from reasoning over long documents and do not trivially generalize to cases where the answer is not present as a span in a given document. We present a novel neural-based architecture that is capable of extracting relevant regions based on a given question-document pair and generating a well-formed answer. To show the effectiveness of our architecture, we conducted several experiments on the recently proposed and challenging RC dataset ?NarrativeQA?. The proposed architecture outperforms state-of-the-art results (Tay et al., 2018) by 12.62\% (ROUGE-L) relative improvement.

    @inproceedings{lirolem34105,
    month = {October},
    year = {2018},
    author = {Satish Indurthi and Seunghak Yu and Seohyun Back and Heriberto Cuayahuitl},
    title = {Cut to the Chase: A Context Zoom-in Network for Reading Comprehension},
    publisher = {Association for Computational Linguistics},
    booktitle = {Empirical Methods in Natural Language Processing (EMNLP)},
    url = {http://eprints.lincoln.ac.uk/34105/},
    keywords = {ARRAY(0x558b5345f730)},
    abstract = {In recent years many deep neural networks have been proposed to solve Reading Comprehension (RC) tasks. Most of these models suffer from reasoning over long documents and do not trivially generalize to cases where the answer is not present as a span in a given document. We present a novel neural-based architecture that is capable of extracting relevant regions based on a given question-document pair and generating a well-formed answer. To show the effectiveness of our architecture, we conducted several experiments on the recently proposed and challenging RC dataset ?NarrativeQA?. The proposed architecture outperforms state-of-the-art results (Tay et al., 2018) by 12.62\% (ROUGE-L) relative improvement.}
    }
  • E. Senft, S. Lemaignan, P. Baxter, and T. Belpaeme, “From evaluating to teaching: rewards and challenges of human control for learning robots,” in Iros 2018 workshop on human/robot in the loop machine learning, 2018.
    [BibTeX] [Abstract] [Download PDF]

    Keeping a human in a robot learning cycle can provide many advantages to improve the learning process. However, most of these improvements are only available when the human teacher is in complete control of the robot?s behaviour, and not just providing feedback. This human control can make the learning process safer, allowing the robot to learn in high-stakes interaction scenarios especially social ones. Furthermore, it allows faster learning as the human guides the robot to the relevant parts of the state space and can provide additional information to the learner. This information can also enable the learning algorithms to learn for wider world representations, thus increasing the generalisability of a deployed system. Additionally, learning from end users improves the precision of the final policy as it can be specifically tailored to many situations. Finally, this progressive teaching might create trust between the learner and the teacher, easing the deployment of the autonomous robot. However, with such control comes a range of challenges. Firstly, the rich communication between the robot and the teacher needs to be handled by an interface, which may require complex features. Secondly, the teacher needs to be embedded within the robot action selection cycle, imposing time constraints, which increases the cognitive load on the teacher. Finally, given a cycle of interaction between the robot and the teacher, any mistakes made by the teacher can be propagated to the robot?s policy. Nevertheless, we are are able to show that empowering the teacher with ways to control a robot?s behaviour has the potential to drastically improve both the learning process (allowing robots to learn in a wider range of environments) and the experience of the teacher.

    @inproceedings{lirolem36200,
    month = {October},
    author = {Emmanuel Senft and Severin Lemaignan and Paul Baxter and Tony Belpaeme},
    title = {From Evaluating to Teaching: Rewards and Challenges of Human Control for Learning Robots},
    year = {2018},
    booktitle = {IROS 2018 Workshop on Human/Robot in the Loop Machine Learning},
    publisher = {Imperial College London},
    url = {http://eprints.lincoln.ac.uk/36200/},
    keywords = {ARRAY(0x558b5393fb28)},
    abstract = {Keeping a human in a robot learning cycle can provide many advantages to improve the learning process. However, most of these improvements are only available when the human teacher is in complete control of the robot?s behaviour, and not just providing feedback. This human control can make the learning process safer, allowing the robot to learn in high-stakes interaction scenarios especially social ones. Furthermore, it allows faster learning as the human guides the robot to the relevant parts of the state space and can provide additional information to the learner. This information can also enable the
    learning algorithms to learn for wider world representations, thus increasing the generalisability of a deployed system. Additionally, learning from end users improves the precision of the final policy as it can be specifically tailored to many situations. Finally, this progressive teaching might create trust between the learner and the teacher, easing the deployment of the autonomous robot. However, with such control comes a range of challenges. Firstly, the rich communication between the robot and the teacher needs to be handled by an interface, which may require complex features. Secondly, the teacher needs to be embedded within the robot action selection cycle, imposing time constraints, which increases the cognitive load on the teacher. Finally, given a cycle of interaction between the robot and the teacher, any mistakes made by the teacher can be propagated to the robot?s policy. Nevertheless, we are are able to show that empowering the teacher with ways to control a robot?s behaviour has the potential to drastically improve both the learning process (allowing robots to learn in a wider range of environments) and the experience of the teacher.}
    }
  • H. Wang, S. Yue, J. Peng, P. Baxter, C. Zhang, and Z. Wang, “A model for detection of angular velocity of image motion based on the temporal tuning of the drosophila,” in Icann 2018, 2018, p. 37–46.
    [BibTeX] [Abstract] [Download PDF]

    We propose a new bio-plausible model based on the visual systems of Drosophila for estimating angular velocity of image motion in insects? eyes. The model implements both preferred direction motion enhancement and non-preferred direction motion suppression which is discovered in Drosophila?s visual neural circuits recently to give a stronger directional selectivity. In addition, the angular velocity detecting model (AVDM) produces a response largely independent of the spatial frequency in grating experiments which enables insects to estimate the flight speed in cluttered environments. This also coincides with the behaviour experiments of honeybee flying through tunnels with stripes of different spatial frequencies.

    @inproceedings{lirolem33104,
    pages = {37--46},
    month = {December},
    publisher = {Springer, Cham},
    booktitle = {ICANN 2018},
    year = {2018},
    title = {A Model for Detection of Angular Velocity of Image Motion Based on the Temporal Tuning of the Drosophila},
    author = {Huatian Wang and Shigang Yue and Jigen Peng and Paul Baxter and Chun Zhang and Zhihua Wang},
    keywords = {ARRAY(0x558b535403a0)},
    abstract = {We propose a new bio-plausible model based on the visual systems of Drosophila for estimating angular velocity of image motion in insects? eyes. The model implements both preferred direction motion enhancement and non-preferred direction motion suppression which is discovered in Drosophila?s visual neural circuits recently to give a stronger directional selectivity. In addition, the angular velocity detecting model (AVDM) produces a response largely independent of the spatial frequency in grating experiments which enables insects to estimate the flight speed in cluttered environments. This also coincides with the behaviour experiments of honeybee flying through tunnels with stripes of different spatial frequencies.},
    url = {http://eprints.lincoln.ac.uk/33104/}
    }
  • H. Wang, J. Peng, and S. Yue, “A directionally selective small target motion detecting visual neural network in cluttered backgrounds,” Ieee transaction on cybernetics, 2018.
    [BibTeX] [Abstract] [Download PDF]

    Discriminating targets moving against a cluttered background is a huge challenge, let alone detecting a target as small as one or a few pixels and tracking it in flight. In the insect’s visual system, a class of specific neurons, called small target motion detectors (STMDs), have been identified as showing exquisite selectivity for small target motion. Some of the STMDs have also demonstrated direction selectivity which means these STMDs respond strongly only to their preferred motion direction. Direction selectivity is an important property of these STMD neurons which could contribute to tracking small targets such as mates in flight. However, little has been done on systematically modeling these directionally selective STMD neurons. In this paper, we propose a directionally selective STMD-based neural network for small target detection in a cluttered background. In the proposed neural network, a new correlation mechanism is introduced for direction selectivity via correlating signals relayed from two pixels. Then, a lateral inhibition mechanism is implemented on the spatial field for size selectivity of the STMD neurons. Finally, a population vector algorithm is used to encode motion direction of small targets. Extensive experiments showed that the proposed neural network not only is in accord with current biological findings, i.e., showing directional preferences, but also worked reliably in detecting small targets against cluttered backgrounds.

    @article{lirolem33420,
    journal = {IEEE Transaction on Cybernetics},
    publisher = {IEEE},
    title = {A Directionally Selective Small Target Motion Detecting Visual Neural Network in Cluttered Backgrounds},
    author = {Hongxin Wang and Jigen Peng and Shigang Yue},
    year = {2018},
    month = {October},
    note = {The final published version of this article can be accessed online at https://ieeexplore.ieee.org/document/8485659},
    abstract = {Discriminating targets moving against a cluttered background is a huge challenge, let alone detecting a target as small as one or a few pixels and tracking it in flight. In the insect's visual system, a class of specific neurons, called small target motion detectors (STMDs), have been identified as showing exquisite selectivity for small target motion. Some of the STMDs have also demonstrated direction selectivity which means these STMDs respond strongly only to their preferred motion direction. Direction selectivity is an important property of these STMD neurons which could contribute to tracking small targets such as mates in flight. However, little has been done on systematically modeling these directionally selective STMD neurons. In this paper, we propose a directionally selective STMD-based neural network for small target detection in a cluttered background. In the proposed neural network, a new correlation mechanism is introduced for direction selectivity via correlating signals relayed from two pixels. Then, a lateral inhibition mechanism is implemented on the spatial field for size selectivity of the STMD neurons. Finally, a population vector algorithm is used to encode motion direction of small targets. Extensive experiments showed that the proposed neural network not only is in accord with current biological findings, i.e., showing directional preferences, but also worked reliably in detecting small targets against cluttered backgrounds.},
    keywords = {ARRAY(0x558b538f3e20)},
    url = {http://eprints.lincoln.ac.uk/33420/}
    }
  • J. Zhao, C. Hu, C. Zhang, Z. Wang, and S. Yue, “A bio-inspired collision detector for small quadcopter,” in 2018 international joint conference on neural networks (ijcnn), 2018, p. 1–7.
    [BibTeX] [Abstract] [Download PDF]

    The sense and avoid capability enables insects to fly versatilely and robustly in dynamic and complex environment. Their biological principles are so practical and efficient that inspired we human imitating them in our flying machines. In this paper, we studied a novel bio-inspired collision detector and its application on a quadcopter. The detector is inspired from Lobula giant movement detector (LGMD) neurons in the locusts, and modeled into an STM32F407 Microcontroller Unit (MCU). Compared to other collision detecting methods applied on quadcopters, we focused on enhancing the collision accuracy in a bio-inspired way that can considerably increase the computing efficiency during an obstacle detecting task even in complex and dynamic environment. We designed the quadcopter’s responding operation to imminent collisions and tested this bio-inspired system in an indoor arena. The observed results from the experiments demonstrated that the LGMD collision detector is feasible to work as a vision module for the quadcopter’s collision avoidance task.

    @inproceedings{lirolem34847,
    booktitle = {2018 International Joint Conference on Neural Networks (IJCNN)},
    publisher = {IEEE},
    author = {Jiannan Zhao and Cheng Hu and Chun Zhang and Zhihua Wang and Shigang Yue},
    title = {A Bio-inspired Collision Detector for Small Quadcopter},
    year = {2018},
    month = {October},
    pages = {1--7},
    url = {http://eprints.lincoln.ac.uk/34847/},
    abstract = {The sense and avoid capability enables insects to fly versatilely and robustly in dynamic and complex environment. Their biological principles are so practical and efficient that inspired we human imitating them in our flying machines. In this paper, we studied a novel bio-inspired collision detector and its application on a quadcopter. The detector is inspired from Lobula giant movement detector (LGMD) neurons in the locusts, and modeled into an STM32F407 Microcontroller Unit (MCU).
    Compared to other collision detecting methods applied on quadcopters, we focused on enhancing the collision accuracy in a bio-inspired way that can considerably increase the computing efficiency during an obstacle detecting task even in complex and dynamic environment. We designed the quadcopter's responding operation to imminent collisions and tested this bio-inspired system in an indoor arena. The observed results from the experiments demonstrated that the LGMD collision detector is feasible to work as a vision module for the quadcopter's collision avoidance task.},
    keywords = {ARRAY(0x558b53877428)}
    }