CAT Framework: Robots Navigate Terrain Based on Their Own Physical Limits

Navigating unstructured terrain is a fundamental challenge for autonomous robots, but a one-size-fits-all approach to traversability fails because the same ground can be safe for a legged robot and deadly for a wheeled one. Existing methods often ignore a robot's unique physical capabilities until late-stage planning, leading to poor decisions. A new paper from researchers led by Gianluca Capezzuto introduces the Capability-Aware Traversability (CAT) framework, which bakes robot-specific constraints directly into the learned spatial representation from the start.

CAT uses Spatially-Adaptive Denormalization (SPADE) blocks to modulate semantic terrain maps with per-robot traversability vectors, and it grounds dense supervision masks in physical trajectories via an interactive annotation pipeline. Tested on human-annotated and trajectory-aligned datasets, CAT achieved an 11.0% improvement in AUROC on physically executed trajectories and a 15.8% boost in AUPRC on human traces. Ablation studies confirm that spatial conditioning and per-robot prototypes create genuine capability sensitivity. The system runs on embedded hardware at 4.8 Hz, and live deployments on a legged quadruped and a wheeled skid-steer demonstrated real-time, embodiment-aware obstacle avoidance. Accepted at IROS 2026, this work moves beyond generic path prediction toward truly adaptive navigation.