EgoRecovery lets robots learn from human failures 10x faster than teleoperation
Scalable robot failure recovery via egocentric human video data—10x more efficient than teleoperation.
A new co-training framework, EgoRecovery, uses egocentric human recovery demonstrations to teach robots how to recover from failures. By aligning human and robot data into a shared corrective-intent space, human operators can generate more than 10x as much valid recovery data per hour compared to robot teleoperation under the proposed protocol. The system improves success from failure starts on real-world recovery tasks over robot-only recovery, direct co-training with human recovery data, and direct intent-transfer baselines.
- Collects recovery data 10x faster per hour than robot teleoperation by using egocentric human videos.
- Aligns human and robot data to a shared 'corrective-intent' space requiring only a small number of robot demonstrations.
- Achieves higher success rates from failure states compared to robot-only and direct intent-transfer baselines in real-world tests.
Why It Matters
Scalable failure recovery is key for reliable real-world robots; this approach dramatically reduces data collection costs.