CARE Paper Turns Robot Execution Failures Into Corrective Demos, Reports 15.9-Point Real-World Dual-Arm Gain
Summary
CARE (Experience-Guided Atomic Corrective Execution) collects failed rollouts from vision-language-action policies, models stage-conditioned post-failure deviations, and synthesizes representative failure states paired with corrective demonstrations. Runtime combines stage-wise planning with 3D monitoring to trigger targeted corrections without discarding task progress. Authors report average gains of 14.5 points in simulation and 15.9 points on real-world dual-arm manipulation.
Originally reported by huggingface.co
Read the original article →Original headline: CARE Paper Turns Robot Execution Failures Into Corrective Demos, Reports 15.9-Point Real-World Dual-Arm Gain