Foundation models are broad AI systems trained on vast text and images. They now serve as robot brains, but these robots can act unsafely, such as dropping electronics in water. Current safeguards miss many physical hazards.
Researchers built RoboRisk, a of over a hundred tasks across five safety categories. They tested six leading models and two robot-specific systems, finding high rates of harmful behavior. They also created LEASH, a safety harness with three gates that review tasks, actions, and irreversible moves using camera images and a shared policy.
LEASH reduced harm in all models, added minimal overhead, and preserved task completion. This shows a simple safety layer can substantially improve robot safety.