Feel-WM is an off-road navigation world model that conditions on proprioception and predicts what the robot will feel (future proprioceptive state and a failure risk) as well as what the camera will see. Both are learned from the robot’s own experience without human labels, and the planner weighs predicted failure risk against goal similarity. It beats visual-only navigation world models in open-loop planning and closed-loop rough-terrain runs on wheeled and legged robots, and on a Husky on mountain trails it completes courses that an end-to-end policy fails.
Relevance
A world model that predicts its own failure risk is a step towards robots that keep out of trouble with no one watching. Read for long-term autonomy and failure recovery.
Links
- arXiv
- Subfield: Robot learning, Safety and verification