Path Tracking for a Skid-steer Vehicle using Learning-based Model Predictive Control

Abstract

Skid-steer vehicle can generate a large traction force, which is especially good for navigation on a rough terrain. However, the turning motion is so sensitive to slippage effect that designing a controller is still a challenging problem. Also, the motion of the vehicle is affected not only by wheel motion, but also by the road properties and the characteristics of wheel control. With this in mind, we employ a model predictive control (MPC) with an on-line model learning. The velocity model, which represents the relationship between true vehicle velocity and input command, is learned with an on-line sparse Gaussian process (GP). The on-line sparse GP can reduce the computational complexity of GP and also consistently update the model from the driving data. Finally, combining with MPC makes it possible to generate an optimal policy based on the learned model. Experiments are conducted to test the tracking performance of a skid-steer robot at the indoor and the outdoor environment. The results show the more reliable performance than the method based on a conventional model with parameter adaptation.