When Black-Box Robot Policies Fail: Online Sensor Diagnosis and Repair with a Transparent Surrogate

Visuomotor policies can execute familiar tasks yet lack the corrective behavior needed after their own mistakes. We present a framework for recursive self-improvement through local recovery supervision. Each round audits the current policy, generates corrective demonstrations at supported failure states, and uses them to update the policy that drives the next round of collection. An offline auditor locates unresolved failures using coarse and dense temporal evidence and specifies observable repair goals. A fixed multimodal agent acts as a tool-using teacher, generating recovery actions through observation, computation, execution, and feedback. The frozen student tests whether each teacher endpoint supports further progress. If continuation fails, the system restores that endpoint and extends the demonstration. Action-level quality assessment then defines continuous training windows with aligned observations, quality weights, and validity masks. Only the student is deployed. In a preliminary LIBERO-Goal study, recovery-augmented post-training achieves 88 successful episodes out of 100 validation scenes, compared with 78 for original-data continuation from the same $π_0$ checkpoint. An earlier BC-RNN study on robomimic Can improves success from 102/130 to 112/130 using 26 local recovery segments. Both comparisons match 2,000 additional optimization steps.

Robot world models support policy evaluation, planning, and synthetic data generation, but these applications require predictions that distinguish successful actions from failures. Across four released checkpoints from two architecture families, we observe weak sensitivity to action changes and success-like predictions on verified failures. Although recent work incorporates failures into model training, which data can repair released checkpoints without changing their architecture or training objective still remains underexplored. To this end, we introduce CureWM, which constructs alternative actions from successful demonstrations across a severity grid, verifies their outcomes through execution in simulation or on hardware, and fine-tunes released models on the resulting failures and surviving successes alongside nominal demonstrations. This construction provides controlled action contrasts from shared starting contexts. On 484 held-out LIBERO failure counterfactuals, optimism falls from 80% after fine-tuning on the official data to 30--43% across four independently fine-tuned CureWM models (38% mean). In two separate evaluations on a physical robot arm, failure predictions scored as success-like by a latent-distance diagnostic decrease from 90% after fine-tuning on successful demonstrations alone to 33% with CureWM. With failure counts per task, successful replay data, and training budget matched, counterfactual failures yield a success--failure value gap of 0.124, compared with 0.014 for freshly collected on-policy failures. These findings support execution-verified counterfactual replay for post-hoc repair and show why reduced optimism must be evaluated alongside success--failure discrimination. Code is available at https://github.com/jiuyixu25/CureWM.

Generalist robot policies carry broad manipulation priors from large-scale data, but specializing them to a new task remains the deployment bottleneck. This requires eliciting task-specific behavior from limited demonstrations without degrading their broad capabilities. We introduce Proxy Policy Steering (PPS), an inference-time adaptation method that resolves this challenge by training two lightweight proxy policies whose calibrated velocity-space difference steers the frozen base sampler. A reference proxy models the frozen base's behavior on target-task observations, and a task proxy, initialized from the reference, captures how this behavior changes under task supervision. Their difference forms a calibrated velocity-space residual that steers the frozen base sampler at every denoising step. We identify the conditions under which this residual isolates the change induced by task supervision, and validate them empirically. Because the base is never directly modified, its broad capabilities remain available at inference, including behaviors such as recovery from failure that the demonstrations themselves do not exercise. Adaptation requires only forward velocity predictions from the base, making PPS lightweight to train and applicable even without access to the base's parameters. On 8 real-world and 4 simulation manipulation tasks, PPS lifts the state-of-the-art pi 0.5 base policy by 53% absolute success rate on average, with zero-to-one gains on tasks the base never solves, while preserving the base's broad capabilities. PPS outperforms LoRA fine-tuning, from-scratch specialists, residual policies, and prior inference-time steering methods.