L-CAS at TAROS 2024
The Lincoln Centre for Autonomous Systems (L-CAS) presented a diverse range of research at the Towards Autonomous Robotic Systems (TAROS) 2024 conference. The contributions spanned various aspects of robotics and autonomous systems, reflecting the multidisciplinary nature of the field.
Presented Research Highlights
- Perrett, A., Brown, J. and Bosilj, P. tackled the challenge of domain shift in deep neural networks, demonstrating a whopping 240% improvement in out-of-distribution performance for ordinal data like age estimation.
- James, K. and Cielniak, G. introduced a novel approach to plant phenotyping, using unsupervised clustering with geometric shape priors to handle occlusions in 3D point cloud data of strawberry stems.
- Heselden, J.R. and Das, G.P. presented “Duckling Platooning,” a safety-first approach to robot navigation in shared spaces, using controlled information disclosure to ensure robust and efficient movement. In platooning approaches, robots are grouped together and planning is completed by the collective.
- Flores, J.P.E., Yilmaz, A., Avendaño, L.A.S., and Cielniak, G. conducted a comparative analysis of Unity and Gazebo simulators for creating digital twins of robotic tomato harvesting scenarios, providing valuable insights for agricultural robotics researchers.
- Zhou, L., et al. developed a 3D printer-based open-source calibration platform for whisker sensors, potentially revolutionizing tactile sensing in robotics.
- Stevenson, R., et al. introduced an open-source hardware whisker sensor, further expanding the toolkit available to robotics researchers and enthusiasts alike.
- Williams, E. and Polydoros, A. investigated pretrained visual representations in reinforcement learning to learn actions from videos.
All papers are available from TAROS’ proceedings, so have a read!
Themes and Implications
The research presented by L-CAS researchers addressed several key themes in contemporary robotics:
- Applied Machine Learning: The work on ordinal regression and unsupervised clustering demonstrates ongoing efforts to refine AI techniques for specific robotics applications.
- Agricultural Robotics: Contributions in plant phenotyping and simulation of harvesting scenarios highlight the growing importance of automation in agriculture.
- Safety and Navigation: The “Duckling Platooning” approach addresses critical issues in robot navigation within shared spaces.
- Simulation and Digital Twins: The comparison of robotics simulators provides valuable guidance for researchers developing digital twins.
- Sensor Technology: The open-source whisker sensor projects contribute to the advancement of tactile sensing in robotics.
These diverse contributions reflect the complex, interdisciplinary nature of modern robotics research. By addressing challenges across multiple domains, from machine learning to hardware development, L-CAS researchers are contributing to the ongoing evolution of robotics and autonomous systems.
The emphasis on open-source hardware and software solutions in several projects aligns with broader trends in the academic community towards open science and collaborative development. This approach has the potential to accelerate progress by enabling wider access to tools and technologies, and stop researchers having to “reinvent the wheel”.