Vision-Language-Action (VLA) policies have become a leading approach for robotic manipulation, but they remain fragile once execution drifts away from the trajectories they were trained on. A new research paper posted on Hugging Face proposes CARE (Corrective Atomic Robotic Execution), a framework designed to make such policies more resilient by learning directly from their failures.
Rather than generating corrective training data through hand-designed or random perturbations, CARE gathers failed rollouts, models how deviations occur after failure at each stage, and uses those empirical distributions to synthesize representative failure states and corrective demonstrations. This data-centric loop aims to close the gap between nominal training conditions and the messy reality of real-world execution.
At inference time, CARE pairs stage-wise planning with physically grounded 3D monitoring. When a deviation is detected, the system can trigger atomic adjustments or re-operations while preserving overall task progress, rather than restarting or abandoning the task.
The authors also introduce the Failure State Recovery Benchmark (FSR-Bench), which measures how well policies recover from intermediate failure states under local deviations and structural anomalies. Experiments spanning multiple VLA backbones, simulation benchmarks, and real-world dual-arm tasks report consistent gains: an average of 14.5 points in simulation and 15.9 points in the real world.
Code, models, and data are available at github.com/xiaojunlan/care. The work sits alongside a growing cluster of 2026 papers on robust robot recovery, including LIBERO-RECOVER, ORPA, and RoboBRIDGE.