CARE: Experience-Guided Atomic Corrective Execution for Vision-Language-Action Policies

Most VLA policies are trained on successful demonstrations and evaluated from clean starting states. The moment something slips, tips, or lands slightly off-target, the policy enters territory it has rarely seen. CARE addresses this by treating those failures as data rather than noise to discard.

The core idea is straightforward: run the policy, collect the rollouts that fail, and characterize how they fail at each stage of the task. Translational and rotational deviations at grasp closure or object release are measured in relative coordinates, then fit to parametric distributions selected by AIC and KS tests. Those fitted distributions are then sampled to synthesize new failure states - not replays of what was collected, but freshly generated physical states drawn from the empirical failure geometry. Corrective demonstrations are collected only for the atomic stage that failed, keeping each corrective trajectory short and targeted.

At inference time, a GPT-4.1 planner decomposes the task once at initialization into atomic stages with associated geometric cues and termination predicates. A 3D monitor - built from SAM 3 masks and Depth Anything 3 depth estimates fused into point clouds - checks geometric conditions at stage-critical transitions. When a deviation is detected, it either issues an intra-execution adjustment or triggers a full re-operation of the current stage. Crucially, the same VLA executor handles nominal, adjustment, and re-operation instructions; there is no separate recovery policy to maintain.

The monitoring comparison is striking. Replacing the point-cloud predicate checker with a GPT-4.1 VLM monitor drops average accuracy from 93.8% to 71.4% and increases latency from 0.24 seconds to 3.40 seconds. Physical grounding matters more than language understanding for this particular decision.

The ablation separates the two contributions cleanly. Corrective training data alone accounts for the larger share of the gain - on the backbone, data alone adds 9.2 points while the execution framework alone adds 4.2 points, but together they add 23.2 points. The interaction is real: corrective supervision without the monitoring mechanism to invoke it leaves significant performance on the table.

Results across simulation benchmarks show average task-success gains of 14.5 points, and real-world dual-arm trials on the SO-101 platform show 15.9 points. Against a reproduced FailSafe baseline on the same hardware, CARE outperforms by 13.7 points for the backbone. The FSR-Bench recovery benchmark - 36 scenarios spanning local geometric deviations and harder structural failures like tipping and target occupation - shows a 7.5-point gain in recovery success rate.

The rollout budget analysis is practically useful: 100 preliminary rollouts stabilize the failure prior well, with Jensen-Shannon divergence dropping sharply between 50 and 100 rollouts and only marginally afterward. That is a concrete number a practitioner can target before investing in corrective data collection.