Stanford AA289/ENGR319 - Robotics and Autonomous Systems Seminar - Stanford Seminar - Leveraging Physics-Based Models To Learn Generalizable Robotic Manipulation

November 8, 2024 Albert Wu, Stanford University Robotic manipulation involves rich contact interactions, which pose significant challenges in producing effective manipulation policies. While model-based planners have achieved great success in continuous systems, their performance typically breaks down when dealing with the complex, discontinuous contact dynamics in manipulation tasks. On the other hand, behavior cloning and reinforcement learning have demonstrated success in fine-grained manipulation, but often struggle with generalization beyond the specific scenarios encountered during training. In this talk, I will explore how incorporating physical models can enable efficient learning of robust manipulation policies. I will present a series of projects that demonstrate how integrating physics priors and contact constraints into learning pipelines can facilitate the execution of sophisticated dexterous and nonprehensile tasks across a wide range of simulation and hardware environments. About the speaker: https://wualbert.github.io/ More about the course can be found here: https://stanfordasl.github.io/robotics_seminar/ View the entire AA289 Stanford Robotics and Autonomous Systems Seminar playlist: https://www.youtube.com/playlist?list=PLoROMvodv4rMeercb-kvGLUrOq4HR6BZD ► Check out the entire catalog of courses and programs available through Stanford Online: https://online.stanford.edu/explore