Agile Off-Road Terrain Traversal of an Ackermann Steered Platform using Deep Reinforcement Learning

Enhancing the mobility and maneuverability of Unmanned Wheeled Ground Vehicles (UGVs) in off-road environments faces many challenges from the diversity of terrain conditions and their engendered uncertainties. Further, the architecture of the wheeled vehicle (skid-steered vs Ackermann-steered) plays a crucial role in developing and successfully executing motion plans. A growing literature showcases the better ability of Ackermann-steered wheeled mobile platforms to regulate their terrain-ground interactions and overall mobility & maneuverability performance in autonomous settings. However, these gains come with requirements to manage the wheel-terrain interactions via more complex planning/navigation to realize the benefits. This is in contrast to the skid-steered platforms that passively accommodate kinematic and dynamic incompatibility via their (uncontrolled) wheel slip. To this end, we present an end-to-end Deep Reinforcement Learning (DRL) framework for an Ackermann-steered wheeled robot in an off-road uneven terrain setting. Within this framework, we explore various facets, multi-objective rewards, varying agent and policy choices to enable better rapid autonomy planning and control of high-speed offroad terrain traversal by Ackermann-steered vehicles. The initial results reported here show promise that the DRL approach can autonomously create dynamically feasible trajectories without explicitly requiring complex user-crafted first-principles modeling.