LeMo · Learning Mobility for Real Legged Robots

Research and applications in legged robotics has made significant progress over the last decade, driven by more advanced actuation systems, better on-board computation, and significantly improved sensors for perceiving the environment. State-of-the-art model-based planning and control algorithms can plan for contact points and body motions to move legged systems over complex environments. However, these methods have shown clear performance limits when it comes to behaviours and situations that are more complex, and it is unclear if and how these limits can be overcome with classical control methods. On the other hand, recent advances in reinforcement learning has put forward unprecedented capabilities to learn control policies for complex behaviours. With our preliminary findings, we were the first group to present methods that allow directly transferring learned behaviours from simulation to reality to create advanced motion skills for complex legged robots. This breakthrough has the potential to revolutionize the field of legged locomotion control. In this ERC project, we want to research this highly promising area and investigate the use of machine learning tools to make legged robots autonomously move in realistic outdoor scenarios. In three parallel research streams, we will learn how to accurately model the system dynamics from experience, how to abstractify, generate and coordinate different complex behaviours that involve multi-contact situations, and how to combine these with perception to enable autonomous navigation in complex environments. The proposed methods have the potential to overcome the limitations of commonly used optimization-based methods such as limited model accuracy, local minima, conservative performance, computational load and execution time, and it will help us to better understand the fundamentals of locomotion in robots and biology.