Reproduce “Adaptive Contact-Implicit Model Predictive Control with Online Residual Learning” on OpenResearch

About this paper

Wei-Cheng Huang, Alp Aydinoglu, Wanxin Jin, Michael Posa

This research introduces an adaptive model predictive control framework for multi-contact robots that incorporates online residual learning to compensate for inaccuracies in hybrid dynamics models, specifically targeting complementarity constraints. The system robustly tracks trajectories for both rigid and non-smooth objects, achieving a 90% initiation success rate for fruit manipulation where unadapted methods completely failed.

Key insights

• Directly adapting the complementarity constraints through residual learning is more crucial and data-efficient for multi-contact tasks than just adapting continuous dynamics, as it targets errors in contact events and forces.
• Integrating online learning directly into a fast, established MPC framework allows for continuous, real-time adaptation to model uncertainties without requiring extensive offline data collection or retraining.
• The use of an Anitescu's LCS formulation for contact dynamics, combined with a QP-based implicit loss function, facilitates relatively fast gradient calculations, enabling the proposed online adaptation rate.

Summary by alphaXiv.