Kai Lu wins two robotics awards at IROS 2024

DPhil student Kai Lu has received two awards at the Environment Dynamics Matters (EDM) workshop during the International Conference on Intelligent Robots and Systems (IROS) 2024 conference, one of the most prestigious events in robotics and embodied AI.     

Kai received the Best Poster Award for the best paper in the workshop. His paper, ‘Learning to Catch Reactive Objects with a Behaviour Predictor’ introduces a new learning-based approach enabling robots to catch moving objects by understanding their reactive behaviours. Unlike predictable objects like a tennis ball, reactive targets adjust their movement in response to the robot, making traditional methods insufficient. The research integrates explicit target state prediction with reinforcement learning, leading to improved anticipatory actions, particularly for evasive targets. This novel methodology has applications ranging from interacting with humans to safely capturing animals for biological research. Kai led the project and was supervised by Andrew Markham, working with fellow student Jiaxing Zhong and alumni Bing Wang and Bo Yang (Hong Kong Polytechnic University).

Kai also received an Excellent Practice Award for his paper ‘Learning Generalisable Manipulation Policy with Adapter-Based Parameter Fine-Tuning’    

This research pioneers the use of adapters in reinforcement learning to enable robotic skill generalisation across multiple robots and tasks without retraining. By learning abstract skills with a virtual manipulator and transferring them to robots with varying physical characteristics, the approach addresses key challenges in adaptability and efficiency.    

Validated through experiments on simulated and real robots, including the Toyota Human Support Robot and ManiSkill, the study achieved impressive results in task success and adaptability. The work was conducted in collaboration with the Dynamic Robot Systems group from the Oxford Robotics Institute (ORI) and offers new pathways for advancing robotic versatility in both academic and practical applications.    

Kai led the project and was supervised by Andrew Markham, working with Ioannis Havoutis, Associate Professor at the Department of Engineering, alumnus Kaichen Zhou, and Kim Tien Ly and William Hebberd of ORI.  

Kai notes that systems like this tackle interesting problems:

Our approach demonstrates an effective and efficient solution to this critical task, showcasing convincing results in both simulation and real robots. The recognition of our work in this prestigious workshop highlights its impact on embodied AI research, particularly in advancing methods for generalisable manipulation across different robotic platforms.

This research looks at two key problems facing real-world robots. Firstly, they have to adapt and react to changes in the world - how can we build a machine reality that allows them to do this? Secondly, once a robotic skill is learnt, how best can we transfer that knowledge to a different type or model of robot? Kai's work pushes these exciting areas forward and has deservedly been recognised with these awards. Andrew Markham, Professor of Computer Science  

About IROS and the EDM Workshop    

The International Conference on Intelligent Robots and Systems (IROS) is one of the largest robotics conferences, spotlighting state-of-the-art developments in the field. The EDM workshop brings together leading researchers and focuses on challenges and solutions for navigating and interacting in dynamic environments, a cornerstone of embodied AI.