Robust and Adaptive Rough Terrain Navigation Through Training in Varied Simulated Dynamics

Sean J Wang, Si Heng Teng, Aaron Johnson

Keywords: rough terrain navigation, model-based reinforcement learning, adaptive control, robust control

Abstract: We propose a model-based reinforcement learning approach for robust and adaptive long-horizon navigation in rough terrain environments. Offline, we train an adaptive dynamics model using a wide range of simulated systems. This model can adapt to any new system using state-transition observations from that system. Predictions from the model capture uncertainty about the system's exact dynamics stemming from insufficient observations. Online, we use a divergence constrained path planner to find routes that are robust to the robot's current understanding of dynamics. In our results, we show this allows for long-horizon driving strategies that are conservative when state-transition observations are limited but have improved performance after giving few state-transition observations.

Submission Number: 14