Learning from Limited Demonstrations: Lessons We Have Learned

Abstract

In this talk, I will explore the role of expert demonstrations in robotics, particularly their dual function as biomimetic benchmarks and accelerators of the learning process. While ideal expert motions are often unattainable, I will discuss three innovative approaches for learning from constrained demonstrations, showcasing their successful application in legged robotics. The presentation will encompass a variety of methods, including generative adversarial imitation learning, unsupervised skill discovery, and the integration of self-supervised models with representation and curriculum learning strategies.