VLA-Corrector: Stage-Aware Observable State Understanding for Prompt-Based Closed-Loop Recovery of Vision-Language-Action Policies

Summary

Abstract Summary

Long-horizon robot manipulation with Vision-Language-Action (VLA) policies remains vulnerable to execution-time deviations, as final task success provides little information for diagnosing and correcting failures caused by action noise, object displacement, or goal misalignment. We introduce a stage-aware failure verification and Prompt Recovery framework that enables closed-loop correction of a fixed VLA policy without parameter updates or privileged simulator states. The framework introduces an observable-history-based Learned Verifier that jointly estimates manipulation progress and execution risk by temporally modeling multi-view visual observations, proprioceptive states, and executed actions. To provide interpretable task understanding, we represent manipulation execution through semantic progress stages, including approach, alignment, grasp, transport, and placement, and identify stage-specific failure patterns. Upon detecting abnormal execution, the framework preserves the original instruction and generates a stage-conditioned recovery prompt, allowing the same frozen VLA policy to produce corrective actions. Extensive multi-round evaluations on LIBERO and LIBERO Plus demonstrate that the proposed approach substantially improves closed-loop reliability under diverse perturbations. Without access to privileged object or goal coordinates, the Learned Verifier achieves recovery performance close to that of the privileged rule-based verifier in the evaluated settings. These results show that observable visual-proprioceptive-action history is sufficient to infer latent task states and enable practical failure recovery for existing VLA policies.

Key Points

Long-horizon robot manipulation with Vision-Language-Action (VLA) policies remains vulnerable to execution-time deviations, as final task success provides little information for diagnosing and correcting failures caused by action noise, object displacement, or goal misalignment.
We introduce a stage-aware failure verification and Prompt Recovery framework that enables closed-loop correction of a fixed VLA policy without parameter updates or privileged simulator states.
The framework introduces an observable-history-based Learned Verifier that jointly estimates manipulation progress and execution risk by temporally modeling multi-view visual observations, proprioceptive states, and executed actions.
To provide interpretable task understanding, we represent manipulation execution through semantic progress stages, including approach, alignment, grasp, transport, and placement, and identify stage-specific failure patterns.
Upon detecting abnormal execution, the framework preserves the original instruction and generates a stage-conditioned recovery prompt, allowing the same frozen VLA policy to produce corrective actions.

Problem Formulation

Let $q$ denote the original task instruction and let the current observation at time $t$ be where $I_{t}^{a}$ and $I_{t}^{w}$ are agent-view and wrist-view RGB images, respectively, and $s_{t}\in\mathbb{R}^{8}$ is the observable proprioceptive state. Let $u_{t}\in\mathbb{R}^{7}$ be the action actually executed in the environment.

The observable history available to the Learned Verifier is The history is temporally ordered: no observation or action from a future time step is used.

Problem Formulation — VLA-Corrector: Stage-Aware Observable State Understanding for Prompt-Based Closed-Loop Recovery of Vision-Language-Action Policies — Figure 1: Overview of VLA-Corrector. The frozen VLA policy interacts with the environment using observable visual, proprioceptive, action, and language histories. A stage-aware verifier estimates execution risk and task stage, after which the trigger gate and recovery adapter generate a stage-specific recovery prompt. The privileged rule-based verifier provides supervision during verifier construction.

Method

The overall architecture of VLA-Corrector is shown in Fig. 1 . The closed-loop correction process is summarized as At each control cycle, the frozen policy outputs an action chunk, yet only a fixed-length prefix is executed before replanning. The verifier subsequently ingests the resulting observable history. If the gating mechanism suppresses an intervention, execution continues under the original task instruction.

Experimental Setup

We evaluate the default OpenPI pi05_libero policy under three benchmark settings: standard LIBERO, LIBERO Plus at difficulty 4, and LIBERO Plus at difficulty 5. Standard LIBERO comprises LIBERO-SPATIAL, LIBERO-GOAL, LIBERO-OBJECT, LIBERO-10, and LIBERO-90. LIBERO Plus adopts the corresponding Spatial, Goal, Object, and 10 suites at difficulty levels 4 and 5.

The three operational configurations are summarized in Table 1 .

Experimental Setup — VLA-Corrector: Stage-Aware Observable State Understanding for Prompt-Based Closed-Loop Recovery of Vision-Language-Action Policies — Figure 3: Success rates in simulation across standard LIBERO and LIBERO Plus at difficulty levels 4 and 5.

Results

We compare three benchmark-level dimensions: standard LIBERO, LIBERO Plus at difficulty 4, and LIBERO Plus at difficulty 5. Standard LIBERO aggregates LIBERO-SPATIAL, LIBERO-GOAL, LIBERO-OBJECT, LIBERO-10, and LIBERO-90. The corresponding success rates are 70.9%, 82.6%, and 80.2% for No Assistance, Rule + Prompt, and Learned Binary + Prompt, respectively.

For LIBERO Plus difficulty 4, the average success rates across the clean and perturbed conditions are 73.7%, 79.7%, and 79.0%.

Results — VLA-Corrector: Stage-Aware Observable State Understanding for Prompt-Based Closed-Loop Recovery of Vision-Language-Action Policies — Figure 4: Online gate diagnostics.

Sources and Demonstrations

Figures are reproduced from Song et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.