Robotic policies learn to recover by training on their own failures
single source· 1 articles · confidence: medium · first seen 2026-09-20 20:00 UTC
What this means for you
If you train manipulation policies, the transferable idea is the data loop: label failures by stage, then reuse them as corrective demonstrations. The code and data are public, so this is cheap to try. Nothing to buy or migrate — the gains are the authors' own, on their own new benchmark, and no evaluation date is given.
A robotics paper proposes training manipulation policies on their own mistakes. CARE collects rollouts in which a vision-language-action policy (a model that maps camera input and an instruction to robot movements) has already gone wrong, groups the failures by task stage, and turns them into corrective training examples, rather than perturbing successful runs at random. At run time, stage-wise planning plus 3D monitoring triggers small adjustments or a re-do without discarding progress. The authors also release FSR-Bench, a benchmark for recovery from mid-task failures. They report average task-success gains of 14.5 points in simulation and 15.9 points on real dual-arm tasks, with code, models and data public; no evaluation date is given.
Key facts
- ·CARE reports average task-success gains of 14.5 points in simulation and 15.9 points on real-world dual-arm tasks, across multiple VLA backbones. source
- ·The method builds corrective data from failed rollouts instead of manually designed or random perturbations. source
- ·The paper introduces FSR-Bench (Failure State Recovery Benchmark), which evaluates recovery from intermediate failure states under local deviations and structural anomalies. source
- ·At inference, CARE combines stage-wise planning with 3D monitoring to trigger atomic adjustments or re-operations. source
- ·Code, models and data are released at github.com/xiaojunlan/care, and the paper gives no evaluation date for the reported gains. source
What the sources say
- Hugging Face Daily Papers (research) — Recovering robot manipulation policies by mining failed rollouts, plus a new recovery benchmark
Sources
The original reporting. Follow these — they did the work.
- Hugging Face Daily PapersCARE: Experience-Guided Atomic Corrective Execution for Vision-Language-Action Policies2026-09-20