Care improves robot action recovery with experience guided corrections
CARE: Experience-Guided Atomic Corrective Execution for Vision-Language-Action Policies
Robotics
Summary
Robots that see and follow instructions can still make mistakes when things don't go as planned. The authors made a system called CARE that learns from these mistakes by studying what went wrong and how to fix it step-by-step. CARE then uses this knowledge to adjust actions during tasks, helping robots recover better and finish their work more successfully. They tested it in simulations and real dual-arm robots, showing it improves task success by about 15 percent.
What this means in practice
- •For robotics engineers: Enable robots to recover from unexpected errors during manipulation tasks by triggering precise corrective actions based on learned failure patterns.
- •For warehouse automation teams: Improve end-to-end success rates for robots handling packages by integrating CARE to correct deviations in pick-and-place operations.
Authors
Junlan Xiao, Junwei Jiang, Zaibin Zhang, Yifan Wang, Zhongbo Zhang, Huchuan Lu, Lijun Wang
Abstract
Vision-Language-Action (VLA) policies achieve strong performance in robotic manipulation but remain brittle once execution deviates from nominal trajectories. We propose CARE (Corrective Atomic Robotic Execution), a framework that improves recovery by learning from failures encountered during execution. Instead of generating corrective data from manually designed or random perturbations, CARE collects failed rollouts, models stage-conditioned post-failure deviations, and uses the resulting empirical distributions to synthesize representative failure states and corrective demonstrations. At inference time, CARE combines stage-wise planning with physically grounded 3D monitoring to trigger atomic adjustments or re-operations while preserving task progress. We further introduce the Failure State Recovery Benchmark (FSR-Bench), which evaluates recovery from intermediate failure states under local deviations and structural anomalies. Experiments across multiple VLA backbones, simulation benchmarks, and real-world dual-arm tasks show consistent improvements, with average task-success gains of 14.5 points in simulation and 15.9 points in the real world. Code, models, and data are available at https://github.com/xiaojunlan/care