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
cs.RO
2026-09-21
CARE fits stage-wise failure distributions from VLA rollouts, synthesizes corrections, and triggers atomic adjust/re-operate via 3D monitoring. π0 rises from 39.6% to 60.7% on hard RoboTwin tasks; real-robot success gains 15.9 points.
Vision-Language-Action (VLA) policies look competent on nominal dual-arm trajectories. Once an object slips, a placement drifts, or two arms fall out of sync, the policy enters states that successful expert data barely covers. Prior recovery work either rolls back and replans, or collects corrections around random perturbations. The first treats failure as an exception to erase. The second rarely matches the errors a given policy actually produces.
CARE, from Dalian University of Technology, asks whether those execution failures can supply the recovery signal.
The first half is data. CARE runs 100 nominal atomic plans without correction and records geometric deviation when a stage misses its termination check: object–gripper misalignment at close for grasps, object–target misalignment at open for placements. For each task and stage it fits Δx, Δy, Δz, and yaw to Gaussian, Beta, Gamma, Weibull, log-normal, and uniform candidates, and keeps the winner by AIC and a KS test. Samples from p(d|k) perturb the relative pose, then physics rolls forward, so the new failure is a generated state rather than a replay. An IK/motion-planning oracle writes short corrective clips in simulation; humans teleoperate on the real robot.
The second half is when to invoke those clips. GPT-4.1 decomposes the task once at the start into atomic stages, each with a nominal instruction, geometric cues, a termination predicate, and candidate adjust / re-operate instructions. No further VLM calls. The monitor is a predicate checker, not a learned judge: SAM 3 masks gripper, object, and target; Depth Anything 3 supplies depth; fused point clouds are tested with skill-level geometric predicates. Consistent geometry keeps the nominal instruction, in-stage drift triggers adjustment, a failed outcome triggers re-operation. The same VLA emits every low-level action.
FSR-Bench separates recovery from nominal success. It holds 36 dual-arm scenarios: 21 easy (missed grasp, placement offset, pose error) and 15 hard (tipping, drop, occupied target). Each episode starts from an intermediate failure.
Simulation methods share 150 nominal demos plus 50 extra trajectories per error type. CARE's extras are experience-guided corrections; baselines get the same count of nominal atomic segments.
| Setting | Baseline | CARE |
| RoboTwin 2.0 hard, π0 | 39.6% | 60.7% |
| RoboTwin 2.0 hard, π0-FAST | 17.4% | 33.1% |
| RoboFactory multi-arm, π0 | 62.0% | 74.0% |
| RoboFactory multi-arm, π0-FAST | 26.7% | 36.0% |
Average task-success gain across benchmark–backbone pairs is 14.5 points. On FSR-Bench, both sides use the same uniformly sampled corrections so the comparison isolates execution. π0 Easy recovery rises from 38.8% to 57.2% and Hard from 15.8% to 21.2%; mean recovery success across three backbones gains 7.5 points.
On a real dual-arm SO-101, four long-horizon tasks at 100 trials each, π0 goes from 30.8% to 50.0% and π0-FAST from 18.8% to 31.3%, beating a full FailSafe reproduction by 13.7 and 7.3 points.
Ablations show data helping more than the execution wrapper: for π0, data-only 31.6% → 40.8%, execution-only 35.8%, both 54.8%. Experience-guided sampling beats uniform random by 12.2 points on π0. The 3D monitor hits 93.8% decision accuracy versus 71.4% for GPT-4.1, at 0.24 s versus 3.40 s per call.
The typical VLA failure is a mid-task deviation, not a blank start. CARE turns the robot's own error distribution into a data engine, and recovery still goes through the original language interface, so there is no extra recovery head. Geometric monitoring is also cheaper and more accurate than asking a VLM whether the grasp worked.
The gain is incremental. Hard-regime π0+CARE still sits at 21.2%. Structural failures remain mostly unsolved.
There is no dedicated limitations section; the gaps sit in the protocol. Failures are modeled as relative translation and yaw, so tipping and drops lean on the execution loop and still recover poorly. A new atomic skill needs a hand-written predicate template and tolerances. The planner is frozen after the opening query. FSR-Bench trains on uniform perturbations and evaluates on empirical failures on purpose, which is fair for the execution ablation and also means the sampling benefit shows up mainly in the standard-task setting. Appendix rollback helps little. Reinforcement learning is named as future work, not done.