Reproduce “Learning Legged MPC with Smooth Neural Surrogates” on OpenResearch

About this paper

Samuel A. Moore, Easop Lee, Boyuan Chen

A framework for learned Model Predictive Control (MPC) for legged robots is presented, leveraging Smooth Neural Surrogates (SNS-MLPs) for dynamics modeling and heavy-tailed Cauchy likelihoods for training. This system enables stable gradient-based MPC, demonstrating robust zero-shot generalization on real hardware for various behaviors and terrains, and achieving up to 2x lower cumulative control cost over sampling-based alternatives.

Key insights

• The quality of derivatives in learned dynamics models, rather than just zeroth-order prediction accuracy (MAE), is paramount for effective and stable control performance.
• Explicitly enforcing neural network smoothness and utilizing robust heavy-tailed likelihoods during training are critical for enabling stable and reliable gradient-based MPC with contact-rich dynamics.
• Gradient-based MPC, when combined with smooth neural surrogates, offers superior computational efficiency, scalability, and reduced hyperparameter sensitivity compared to sampling-based MPC for complex legged robot tasks.

Summary by alphaXiv.