CARE: Experience-Guided Atomic Corrective Execution for Vision-Language-Action Policies
CARE trains recovery behavior from the robot’s own failed rollouts, then uses it to correct mid-task deviations.
The framework models post-failure states by task stage and synthesizes corrective demonstrations from those empirical failure patterns. At inference, it monitors execution in 3D and triggers small atomic fixes or re-operations without discarding progress. The authors also introduce FSR-Bench for testing recovery from intermediate failures. They report average task-success gains of 14.5 points in simulation and 15.9 points on real dual-arm tasks. HF Daily Papers' note
The framework models post-failure states by task stage and synthesizes corrective demonstrations from those empirical failure patterns. At inference, it monitors execution in 3D and triggers small atomic fixes or re-operations without discarding progress. The authors also introduce FSR-Bench for testing recovery from intermediate failures. They report average task-success gains of 14.5 points in simulation and 15.9 points on real dual-arm tasks. HF Daily Papers' note
score 5