Sampling-based Model Predictive Control With CVaR-based Risk Assessment for Off-road Environments

Autonomous off-road navigation is challenging due to the presence of unstructured terrain and environmental uncertainty. We propose a sampling-based model predictive path integral control method with terrain-aware risk assessment using conditional value at risk (CVaR). This approach enables robust path planning without requiring differentiable cost functions or reference trajectories. Elevation-based maps are utilized to generate a CVaR risk map, allowing conservative planning through recursive risk evaluation. Simulations and real-world tests using a ground robot demonstrate an improved success rate, smoother motion, and safer navigation compared with hybrid A* and model predictive control methods.