Rough Terrain Path Tracking of an Ackermann Steered Platform using Hybrid Deep Reinforcement Learning

Accurate path tracking of wheeled Uncrewed Ground Vehicles (UGVs) in off-road environments faces numerous challenges stemming from the diversity of operational conditions. Traditional model-based controllers for Ackermann-steered vehicles feature good (skid-free) path-tracking performance on flat ground, but performance degrades with increasingly uneven terrain and faster traversal speeds. This paper introduces a novel approach, a Hybrid Deep Reinforcement Learning (HDRL) controller, leveraging the strengths of a Linear Quadratic Regulator (LQR) and a Deep Reinforcement Learning (DRL) controller, for the enhanced path tracking of Ackermann-steered UGVs. The DRL controller primarily compensates for uncertainty in terrain conditions and unknown vehicle parameters but can be computationally expensive to train. The LQR controller guides the DRL controller during the initial training phases, ensuring a more stable performance and achieving higher rewards in early iterations. In doing so, this hybrid methodology offers promise to overcome the limitations of model-based controllers and the sample-inefficient nature of conventional DRL approaches. Preliminary results showcased in the manuscript show promise for the HDRL controller, demonstrating better performance than model-free DRL and conventional feedback controllers.