Learning-Based Model Predictive Control for Quadruped Locomotion on Slippery Ground

Abstract

Nowadays, reinforcement learning (RL) and model predictive control (MPC) are two of the most widely used methods in robotics community. Model-based MPC enable the robot with stable locomotion capabilities, while Model-free RL provide an automatic approach to learn the policy to maximization the corresponding task performance. In this work, be aiming at utilize the advantages of these two approaches, we propose a Learning-Based Model Predictive Control (LBMPC) methodology for quadruped robot which improves MPC performance by learning the upper-layer decision parameters for MPC though a Heuristic Monte-Carlo Expectation-Maximization (HMCEM) algorithm. We validate this framework with the problem of dynamic locomotion on slippery ground by learning the friction factor which be fixed in standard MPC algorithm. Simulation results show that our LBMPC succeeds in find the optimal friction factor respect to different ground, and our heuristic overcome the problem that the conventional EM algorithms is sensitive to the initial value of policy. At last, we deduce a heuristic strategy for crude but fast ground classification based on empirical data.