Task-Conditioned Uncertainty Costmaps for Legged Locomotion

Legged robots rely on precise foot placement to maintain dynamic stability, especially on uneven or unknown terrain. Traditional geometric costmaps often fail when the terrain differs from training data, leading to planning errors. In a new paper, Singh et al. introduce Task-Conditioned Uncertainty Costmaps that model epistemic uncertainty in predicted footholds. By conditioning on both terrain observations and commanded motion, the model identifies when it is operating in out-of-distribution (OOD) regimes—terrain not seen during training. This uncertainty is then integrated into a unified costmap, guiding the robot away from risky areas where its predictions are unreliable.

In experiments, the approach achieved a 37% reduction in feasibility error in simulation compared to geometry-only baselines, with improved OOD detection and more reliable planning behavior on both simulated and real-world unstructured terrain. The key insight is that a single learned model, even when trained on limited data, can express uncertainty caused by missing training coverage, enabling safer autonomous navigation. This work directly addresses a critical bottleneck in deploying legged robots in search-and-rescue, exploration, and other environments where terrain unpredictability is the norm.