Can robots learn what ground is safe to walk on?

Robots struggle to safely navigate unfamiliar terrain because existing methods either rely on hand-coded rules that don't transfer across platforms or require expensive labeled data. COTRATE learns traversability (whether terrain is safe/efficient) directly from a robot's own unlabeled experience—sensors like accelerometers and cameras—while continuously updating online. A diversity-aware memory strategy keeps computational overhead low for onboard deployment. Tested on 50,000 images across 11 terrains and three environments, the approach transfers knowledge between robots with different body designs. Code and dataset released.