Value-Guided MPPI for Off-Road Navigation

Safe, high-speed autonomous driving over unstructured off-road terrain is difficult: terrain geometrics like steep slopes can cause unsafe driving conditions, and building accurate, explicit cost maps of the terrain is expensive. This work develops a hierarchical framework in which we use MPPI as a high-level waypoint path planner and an RL policy as a low-level controller to track these waypoints. We include the RL policy’s value function as an execution-aware planning cost inside Model Predictive Path Integral (MPPI) control, enabling bidirectional feedback between planning and control without explicit terrain cost maps (Alavilli et al., 2026).

We evaluate our results in the BeamNG simulator, demonstrating 86% navigation success versus 55% for baseline MPPI across unseen off-road evaluation tasks, while substantially reducing vehicle jerk, acceleration, pitch, and roll. We also use online adaptation to update the value function with driving experience.

This work was accepted as a full paper with oral presentation at IEEE IROS 2026. Arxiv link coming soon!

Since publishing this work, we have deployed the algorithm on a full-size Yamaha ATV in collaboration with the Airlab at CMU. Stayed tuned for sim-to-real work on this!

Download Poster (PDF)