acceptodds
Under review as a conference paper at ICLR 2027

CASH: Calibrated Self-Healing Vision-Language-Action Policies for Embodied Agents

Abstract

Action-chunking vision-language-action (VLA) policies predict multiple future actions and execute them without incorporating new observations, local execution errors can thus accumulate over the open-loop horizon. To address this issue, we present CASH (Calibrated Self-Healing), a lightweight failure-detection and self-healing framework that keeps the VLA frozen and consists of detection and self-healing. The detection module contains an execution anomaly detector (EAD) and a policy anomaly detector (PAD). EAD tests whether an executed action is consistent with the observed latent transition. Its score combines forward prediction error, directional disagreement, and inverse-dynamics reconstruction error; a conformalized quantile threshold with a one-sided CUSUM (cumulative sum) that converts the score into a state-conditioned trigger for persistent deviations. PAD tests whether a proposed action is plausible under the current observation and instruction, using successful actions and several families of counterfactual actions to train a value model. When EAD triggers, CASH computes the latent displacement that should have occurred but has not yet occurred, and adjusts the action chunk through a lightweight dynamics model. When PAD triggers, CASH shortens the action chunk and triggers earlier re-planning; repeated low scores activate a window reset for severe out-of-distribution (OOD) conditions. The detection modules are trained on cached latent features using successful trajectories only, without failure data or manual failure labels. The self-healing module does not require additional training. CASH can be integrated into different VLA models without retraining the VLA backbone, while preserving the efficiency of action chunking.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot self-healing, Vision-Language-Action model, Anomaly detection, Conformal prediction, Failure recovery

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_hCYgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CASH, a framework for detecting and recovering from failures of frozen action-chunking VLA policies. The authors distinguish two types of failure: execution failures, in which the executed actions do not produce the expected state transition, and policy failures, in which the proposed action plan itself is wrong. Following this motivation, the authors propose two modules, the Execution Anomaly Detector (EAD) and the Policy Anomaly Detector (PAD). EAD checks whether the observed latent transition is consistent with the change that the executed actions should have produced. PAD is a value model that judges, before execution, whether a proposed action chunk is plausible under the current state and instruction. EAD is built on a latent forward–inverse dynamics model trained with cycle consistency, and both detectors are trained only on successful trajectories. Through experiments, the authors show that the proposed detection and recovery improve success rates, and that the detection modules detect failures better than existing baselines.

Critical strengths
  1. The paper tackles an important yet challenging problem: recovering from failures without failure data or recovery demonstrations.
  2. The motivation for decomposing failure handling into two separate modules is clear.
  3. The experiments support several of the main claims. The detection module raises alarms earlier than the baselines and achieves higher failure recall, and its advantage remains after the length bias is removed (AUC_H4 in Table 2). Using both modules together yields the highest success rate in Table 3. Figure 3 also shows why the inverse head is needed, as it increases the action sensitivity of the forward model.
Critical weaknesses
  1. Heuristic recovery strategy. The failure detection method itself is reasonable, and decomposing it into two modules is a good design. However, the recovery procedures, in particular the PAD path (foreshortening to N_v steps and a window reset triggered by a fixed event count), rely on hand-designed rules. The paper would benefit from further explanation of these design choices.

  2. Readability of Figure 1. Figure 1 is difficult to read. Text elements overlap and the overall layout is complex, so the figure needs improvement.

  3. Link between action sensitivity and performance. Figure 3 shows that the inverse head increases the action sensitivity of the forward model. However, the paper does not show whether this improvement translates into better detection or recovery, for example by reporting the detection AUC or success rate of the forward-only variant.

  4. Contribution of the privileged reset. In Table 3, the PAD configuration combines foreshortening and the window reset. The reset is privileged: it writes the initial joint state directly to the simulator, and its steps are not counted against the episode budget. It is therefore unclear how much of the +9.5 pp gain from PAD comes from the reset, and no ablation separates the two.

  5. Lack of qualitative results. The paper provides no videos or snapshots of the recovery process, which makes it difficult to see how the method operates in practice.

Questions for authors
  1. The recovery strategy, especially the PAD path, appears heuristic overall. Could the authors explain the reasoning behind these design choices?

  2. I suggest improving the readability of Figure 1.

  3. Could the authors provide an ablation that separates foreshortening and the window reset in the PAD path, so that the contribution of the privileged reset is clear?

  4. Could the authors provide qualitative results, such as videos or snapshot sequences of the recovery process?

  5. The title refers to “embodied agents,” but the method and experiments focus on runtime monitoring and recovery for action-chunking VLA policies in simulated tabletop manipulation. I suggest revising the title, or clarifying in the paper, so that the stated scope matches the evaluated setting.

Reviewer_ouwjOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes CASH, a failure-detection and recovery framework for action-chunking VLA policies. The VLA is kept frozen, and two detectors trained only on successful trajectories are attached: an execution anomaly detector (EAD), which uses forward and inverse dynamics models to check whether an executed action is consistent with the observed latent transition, and a policy anomaly detector (PAD), which uses a value model to judge whether a proposed action is plausible given the observation and instruction. When EAD triggers, the action is corrected by gradient descent through the forward model; when PAD triggers, the execution horizon is shortened, with a window reset after repeated triggers. On Meta-World and LIBERO, CASH reports higher success rates than the base VLAs and VLA-Corrector.

Critical strengths

The problem is important. Action-chunking VLA policies execute several actions without new observations, so execution errors can accumulate, and runtime failure detection and recovery are essential for reliable deployment. The setting of an add-on monitoring and recovery layer that keeps the VLA frozen and is trained only on successful trajectories is practically well motivated. In addition, the analysis in Figure 4, showing that PAD triggers concentrate on states that never appear in the demonstrations ("gripper closed but not holding the object"), offers a physically interpretable account of what the detector captures.

Critical weaknesses
  1. The source of the contribution is unclear. The overall structure of CASH closely follows existing work. Monitoring with a latent forward model, truncating the remaining chunk, and correcting it with gradients is shared with VLA-Corrector, which the paper uses as its baseline. Anomaly detection combining forward and inverse dynamics is a classical design, and conformalized quantile regression and CUSUM are standard tools. Splitting failures into execution failures and policy failures also has precedent (e.g., Sentinel). The paper combines many components, but it does not make clear whether the contribution is a new idea specific to this problem, the combination itself and its engineering performance, or a new technical element. The title and Figure 1 likewise list components without conveying a central idea.

    The strongest candidate for a new technical element is the counterfactual cycle-consistency term. However, the problem it addresses, namely that a forward model trained on expert data can ignore the action, is well known (cf. ICM, Pathak et al., 2017; causal confusion, de Haan et al., 2019; counterfactual data augmentation, Pitis et al., 2020). The paper does not discuss its relation to this literature, and the effect of the term is shown only through the sensitivity metrics in Figure 3; its contribution to detection performance or task success is not demonstrated.

  2. The figures and tables fall below the standard expected at ICLR. Figure 1 crowds every component into the figure without conveying the main point, and much of its text is hard to read. In Figure 3, the text is extremely small, the table on the right is embedded as an image and is inconsistent with the main text (RrankR_{\text{rank}} = 0.595 in the figure vs. 0.596 in the text), and the three curves largely overlap, making the claimed difference hard to see. In Figure 5, the legends overlap the plotted curves. Table 1 does not highlight the best results, and Table 2 has too many columns to read easily. These issues directly hinder reading the results.

Questions for authors

Clarifications (could change my assessment)

  1. Window reset. According to Appendix C, the reset writes the arm joint state directly back to its initial value in the simulator, and the reset steps and subsequent re-inference are not counted against the episode step budget. This is a privileged operation that cannot be performed on a real robot, and it effectively extends the time limit. Please report:

    • Tables 1 and 3 without the reset;
    • results where the reset is executed as physical actions and counted against the step budget.
  2. Isolating the source of the gains. PAD's response amounts to executing fewer steps and re-planning, and it increases the number of VLA calls (Table 3). Please compare against simple baselines:

    • always using a shorter execution horizon;
    • random re-planning with a matched number of VLA calls.

    For EAD, please compare the gradient-based correction with simply re-querying the VLA upon a trigger. Comparisons with methods that use new observations during chunk execution (e.g., RTC, or arXiv:2509.23224, which the paper cites) would also help.

Suggestions

  1. Please state the central contribution explicitly, and position the cycle-consistency term within the literature on action-insensitive dynamics and world models.
Reviewer_iZWAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CASH, which uses two additional modules to detect and correct failures during execution without modifying the VLA weights. The first module, the execution anomaly detector (EAD), performs checks after actions are executed. It uses a forward model to predict how the preceding action segment should have changed the visual features, then compares the predicted change with the actual change. When the discrepancy is too large, it uses gradient descent to adjust the next action segment to compensate for the expected change that did not occur. The second module, the policy anomaly detector (PAD), performs checks before actions are executed. It uses a value model to predict whether an action segment is reasonable given the current image and instruction. When the model predicts that the action is unreasonable, PAD executes only the first few steps before running inference again. If PAD is triggered repeatedly, it resets the arm to its initial pose. Both modules require only successful demonstrations for training. Experiments on Meta-World and LIBERO show improvements with SmolVLA, π0, and π0.5.

Critical strengths

The method achieves improvements while keeping the VLA frozen, and it improves performance across three different VLA backbones. This suggests that it could be directly integrated with different VLAs. The two lightweight modules also require only successful demonstrations for training and do not require additional failure data, so the additional overhead is limited.

Critical weaknesses

EAD uses the state expected by the old plan as its correction target. If the old plan is itself flawed, this correction may instead pull the robot back toward an incorrect plan. The correction also relies entirely on a forward model trained only on successful demonstrations, and this model may not make accurate predictions in OOD states. The paper does not validate this issue. The experiments also do not include disturbances such as pushes or slips. The execution errors that EAD is intended to detect are therefore never tested directly, so there is still insufficient evidence to support the effectiveness of this correction method.

Questions for authors
  1. A reset directly returns the arm to its initial pose, and the reset steps do not count toward the episode step limit. This effectively provides an additional retry opportunity. How many times was reset actually triggered in the experiments, and what would the results be if reset were disabled?
  2. When PAD is triggered, it executes only the first few steps before running inference again. This effectively shortens the length of each execution segment. What would the results be if the policy always executed shorter segments, with the same number of VLA calls as CASH, or if triggering occurred randomly at the same frequency?
  3. The VLA already generates new actions based on the current image. Why does EAD still attempt to return to the state expected by the old plan? How do the two approaches compare? Has the prediction accuracy of the forward model in OOD states been validated?
  4. The calibration guarantee in the paper applies only to a single frame, while actual deployment also introduces a CUSUM threshold. Does the calibration data come from demonstrations or rollouts in the simulator? What is the false trigger rate on successful episodes when using the deployment trigger parameters, including CUSUM?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.