Reinforcement Learning Based Online Parameter Adaptation for Model Predictive Tracking Control Under Slippery Condition

Huidong Gao, Rui Zhou, Masayoshi Tomizuka, Zhuo Xu

Abstract: Wheeled mobile robots have a great variety of applications both indoors and outdoors. They are often required to work in various environments such as on rough terrains, wet roads, icy roads, and routes with rapid cornering. Therefore, it is important to design a control scheme that delivers robust tracking performance on various terrains, especially the ones that induce skidding and slipping easily. In this work, we focus on the challenge of mobile robot motion under slippery conditions, and propose a hierarchical framework that learns to adapt the control parameters online for robust model predictive tracking control of mobile robots. Concretely, the high level module based on reinforcement learning (RL) actively adjusts the model predictive control (MPC) scheme to be more aggressive or conservative, by adapting key parameters in the MPC optimization formulation. Experiments demonstrate our framework achieves adaptive and robust tracking performance, especially at rejecting slipping and reducing tracking errors when the mobile robot travels through various terrains. Our framework neither relies on knowing specific dynamic parameters nor requires data fitting. The model trained in simulation is validated in a zero-shot manner with completely different real-world terrain conditions to demonstrate its adaptability.