Explore the cutting-edge research on improving physics simulation for AI applications in this 59-minute Stanford University seminar. Delve into Professor Karen Liu's work on overcoming the sim-to-reality gap by enhancing physics engines rather than control policies. Learn about the development of "learnable" physics engines, efficient training techniques, and progress in sim-to-real transfer involving human interaction. Gain insights into topics such as torque limits, gradient computation, human-aware robust sensing, and challenges in nonlinear dynamics. Discover how this research impacts the safe learning of robots in physical human-robot interaction scenarios without risking real people.
Overview
Syllabus
Introduction
Physics Engine
Evaluation
Learning Opportunities
Torque Limits
Gradient Computation
HumanAware Robust Sensing
Questions
Is your code available
Data collection policy
Stability issues
Nonlinear and discontinuous dynamics
Uncertainty in the state
Taught by
Stanford HAI
