GPS-Denied Rover Motion Regulation as a Physically Grounded POMDP for Recurrent Reinforcement Learning

Saki Omi, Miguel Olivares-Mendez

Keywords: Reinforcement Learning for POMDPs, Planetary robotics, Physically grounded POMDP, Latent disturbance, Action history

TL;DR: We formulate GPS-denied rover motion regulation as a physically grounded POMDP to evaluate recurrent RL agents under hidden wheel slip and gyroscope bias.

Abstract: We present a physically grounded engineering case study for evaluating recurrent reinforcement learning (RL) under partial observability in space robotics. The scenario models a wheeled rover performing motion regulation in a GPS-denied environment, where wheel slip and gyroscope bias corrupt onboard sensor readings, yielding a Partially Observable Markov Decision Process (POMDP). Unlike abstract POMDP benchmarks, the latent disturbances are grounded in established sensor error models and terrain mechanics, making the problem directly relevant to planetary exploration. We propose this scenario as a benchmark primitive for latent-disturbance inference — not a complete navigation solution — and formulate a hypothesis-driven experimental design to test whether action-history inclusion improves inference of hidden disturbances and whether training under diverse abstract disturbances transfers to this physically grounded setting.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 17