RESETTLERobotic Recovery through Disagreement-Triggered Retrieval and Efficient Corrective Control

Yuxin Chen1,*Senqiao Yang2,*Zixuan Wang1Jinhui Ye1Changsheng Lu1
Pengguang Chen3Shu Liu3Zhuotao Tian4,5Jiaya Jia1
1The Hong Kong University of Science and Technology 2The Chinese University of Hong Kong 3SmartMore 4Harbin Institute of Technology, Shenzhen 5Shenzhen Loop Area Institute

* Equal contribution.

Detect disagreement. Retrieve experience. Recover and resume.

A capable robot policy can still lose progress after an imperfect grasp, an unexpected contact, or a shifted initial state. RESETTLE adds a recovery layer at the action-execution interface. It detects persistent disagreement between independently sampled action proposals, retrieves a successful same-task demonstration, and executes one reference-guided corrective action before returning control to the base policy.

The base policy stays frozen. Recovery uses a state-servo prior and a guarded visual residual, without additional vision-language reasoning or online trajectory optimization.

See RESETTLE in action.

+8.70 ppLIBERO-PlusFastWAM · aggregate success
+6.28 ppMeta-WorldSmolVLA · difficulty-group average
+6.83 ppRoboCasa TabletopQwenGR00T · task-average success

Detect. Retrieve. Correct.

Successful demonstrations provide a reference for local recovery. An adapted V-JEPA encoder connects the current observation to these references, and a lightweight controller translates the match into a corrective action.

RESETTLE framework: persistent base-policy disagreement triggers V-JEPA retrieval; the retrieved image and robot state guide state-servo and guarded visual-residual control.

Wait for persistent disagreement.

Sample two action chunks under identical conditioning. Intervene when a training-calibrated disagreement threshold is exceeded at two consecutive decisions.

Find a same-task reference.

Match the current visual representation to successful demonstrations. Retrieve the reference observation together with its robot state.

Make one correction. Resume.

Refine a state-servo prior with a bounded, direction-checked visual residual. Execute one native action, re-observe, and return control to the policy.

Why action disagreement?

Under the same observation and instruction, unstable action choices can signal that a policy needs help. In QwenPI rollouts on LIBERO-Plus, failures show heavier disagreement tails than training-data inference and successful execution.

Action disagreement histograms for layout and robot-initialization perturbations on LIBERO-Plus, comparing training data, successful rollouts, and failed rollouts.

Simulation across six base policies.

RESETTLE improves aggregate results on three benchmarks while keeping the original policy weights fixed. Gains are reported in percentage points (pp).

LIBERO-Plus aggregate success rate (%)
Base policyWithout recovery+ RESETTLEGain
FastWAM49.358.0+8.70 pp
π0.570.073.7+3.70 pp
QwenPI80.284.6+4.40 pp
Experimental settings across LIBERO-Plus, LIBERO-Pro, Meta-World, RoboCasa, and real-world Franka manipulation.

Efficient corrective control.

83.6 ms

Recovery computation
with QwenPI

In the measured QwenPI implementations, RESETTLE reduces monitoring-and-recovery computation by 74.04–93.57% relative to VoLoAgent’s place and grasp branches. RESETTLE produces one action; VoLoAgent plans a trajectory.

Timings cover computation only, excluding nominal policy inference, polling delays, and robot execution.

RESETTLE
83.6 ms
VoLoAgent · place
322 ms
VoLoAgent · grasp
1,300 ms

Local recovery also helps an agentic planner.

Within Harness VLA, RESETTLE improves LIBERO-Pro Swap success with the same GPT-5.5 planner and frozen π0.5 policy, evaluated on the same 100 trials.

42%50%

Recovery beyond simulation.

On a Franka Research 3, one recovery module supports both QwenPI and VLAct across button pressing, cube stacking, object placement, and spoon use. The study includes 320 trials: 20 trials per task for each of four policy configurations. Spoon use measures completion of three stages; the other tasks measure success. The final column averages these four task metrics.

Real-world QwenPI plus RESETTLE rollouts on a Franka arm: placing a carrot in a pot, pressing a button, stacking cubes, and using a spoon.
Real-world task metrics (%)
PolicyPress button
Success
Stack cubes
Success
Carrot in pot
Success
Use spoon
Stage completion
Mean task
metric
QwenPI90558536.766.67
+ RESETTLE100709550.078.75
VLAct90859080.086.25
+ RESETTLE100909586.792.92

Explore the code.

The open-source release provides the LIBERO reference implementation: joint recovery-model training, expert goal-bank encoding, an HTTP recovery service, and a guarded single-action executor.

Citation.

@article{chen2026resettle,
  title={RESETTLE: Robotic Recovery through Disagreement-Triggered Retrieval and Efficient Corrective Control},
  author={Chen, Yuxin and Yang, Senqiao and Wang, Zixuan and Ye, Jinhui and Lu, Changsheng and Chen, Pengguang and Liu, Shu and Tian, Zhuotao and Jia, Jiaya},
  journal={arXiv preprint arXiv:2610.12185},
  year={2026}
}