Improving Trajectory Tracking Accuracy for Faster and Safer Autonomous Navigation of Ground Vehicles in Off-Road Settings

Abstract

Autonomous navigation in off-road settings is complicated by environment-induced disturbances due to natural phenomena, such as friction and slip. This introduces deviations in trajectory-following capabilities, exacerbates the effects of inevitable understeering and, in the worst case, can lead to unsafe navigation. We anticipate that future systems will need to learn terrain-induced effects efficiently from experiences on-platform, in real-time, to compensate for complex terrain. Toward this vision, we discuss a data-driven approach to improving trajectory tracking accuracy by combining conventional, model-based solutions that use proprioceptive sensor data with online, self-supervised learning of noisy disturbances using exteroceptive data. We investigate the value and challenges of predicting disturbances and computing corresponding command offsets based on the robot's experiences.