CARE: Experience-Guided Atomic Corrective Execution for Vision-Language-Action Policies
Junlan Xiao, Junwei Jiang, Zaibin Zhang, Yifan Wang, Zhongbo Zhang, Huchuan Lu, Lijun Wang
Dalian University of Technology
Published September 21, 2026 · on the feed September 22, 2026 · ▲ 24 · arXiv 2609.24118
The 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
The Brief
2:04A short podcast summary of the paper. Written and voiced by AI.
Transcript
CARE. Experience-Guided Atomic Corrective Execution for Vision-Language-Action Policies. By Junlan Xiao and six others, from the Dalian University of Technology. Published on the twenty-first of September. Its code repository is worked out: about one thousand source files, with a readme and a licence.
Vision-language-action policies perform well at robotic manipulation, as long as things go to plan. Once execution drifts off the nominal trajectory, they become brittle. One way to teach recovery is to generate corrective data from perturbations that are designed by hand or chosen at random. This work takes the other route.
CARE, short for Corrective Atomic Robotic Execution, learns recovery from the failures that really happen during execution. It collects failed rollouts, and models the deviations that follow a failure, conditioned on the stage of the task. The empirical distributions from that modeling are used to synthesize representative failure states, along with demonstrations of how to correct them.
At inference time, CARE pairs stage-wise planning with 3D monitoring that is physically grounded. When the monitoring warrants it, CARE triggers either a small atomic adjustment or a re-operation, and does so while preserving the progress the task has already made. Recovery does not mean starting over.
To measure this ability, the authors also introduce the Failure State Recovery Benchmark, FSR-Bench. It evaluates recovery from intermediate failure states, under both local deviations and structural anomalies.
Experiments span multiple VLA backbones, simulation benchmarks, and dual-arm tasks in the real world, and the improvements are consistent. Average task success rises by fourteen point five points in simulation and by fifteen point nine points in the real world.
CARE's contrast is between imagined failures and observed ones. Rather than guessing how a robot might go wrong, it records how it does go wrong, stage by stage, and trains correction on that. Then, when a failure comes, the fix is sized to it: an atomic adjustment where that is enough, a re-operation where it is not, and the progress already made is kept either way.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A robot that only knows clean runs is one slip away from failure. Here is the wall.
You run a robot arm policy trained on clean, successful demonstrations. It works well when every episode starts the way the demonstrations did.
The road ahead
- 1 · The problem ~7 min A robot policy trained only on clean, successful demonstrations breaks down once a slip pushes it into states it has rarely seen.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Treat the policy's own failures as data: model how it fails at each stage, train on fixes for those failures, and monitor each stage to apply them.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Collect failures, model them per stage, synthesize failure states with short fixes, then monitor each stage and adjust or redo it.
- 6 · The experiment ~6 min You’ll guess: On real-world dual-arm tasks, how many points did this failure-driven recovery method add to average task success?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.