Learning Based Controllers for Off-Road Trajectory Tracking of Unmanned Ground Vehicles

Abstract

Vehicles operating in off-road environments face various challenging scenarios, robust operation in these settings requires the trajectory tracking controller to deal with interaction dynamics of specific off-road vehicles with different types of terrains. Tire model based approaches try to solve this issue by explicitly modeling these complex interactions, but they require the measurements and estimations of numerous parameters using specific sensor setups which make them difficult to be used on a generic off-road vehicle. A learning based approach that can bypass the requirement of explicitly modeling these interactions, be able to generalize to different off-road vehicles and capable of life-long learning and adaptation is an attractive candidate to solve off-road trajectory tracking problem. In this work, we first developed a model-based approach that avoids the requirement of explicitly modeling the tire-terrain interactions by using human-driven trajectories to learn the vehicle model. This approach solves the modeling issue but uses a linear approximation of the model and online optimization, which is computationally expensive. Furthermore, it involves online tuning of controller parameters on real vehicles, which requires significant time, domain knowledge, and effort. To deal with these issues, we developed a second model-free approach that learns in simulation and adapts to the real vehicle by utilizing demonstration data. The proposed method avoids online optimization and runs 15 times faster than the model-based approach. Furthermore, it does not require online tuning of controller parameters, significantly reducing the time and effort to implement it on a real vehicle. We also show that this method can seamlessly generalize to different off-road vehicles with similar kinematics but significantly different dynamics without online parameter tuning and only utilizing demonstration data.