acceptodds
Under review as a conference paper at ICLR 2027

PolicyGround: Recovering from High-Level Failures by Probing Frozen Hierarchical Vision-Language-Action Models

Abstract

Hierarchical vision-language-action (VLA) systems support long-horizon manipulation by decomposing a task into language subtasks executed by a low-level policy. After a high-level failure, however, a semantically correct recovery instruction may still be unsupported by the particular policy that must execute it. Assessing this support is difficult for multi-step recovery: later policy responses depend on states produced by earlier actions, so policy support cannot generally be assessed from a single policy query, while physically trying each candidate defeats one-shot recovery. We introduce \method, a training-free framework for recovering from high-level failures in frozen, black-box hierarchical VLAs. \method first derives a policy-independent recovery goal from the recorded failed episode. Its key mechanism, Chain-of-Probe, then connects high- and low-level target-policy responses through proxy contexts retrieved from the failed episode, constructing a proxy multi-step response before physical execution. This policy-specific evidence is used to evaluate and revise the recovery instruction toward one supported by the target policy. We evaluate \method with RoboMME-FR, a one-shot physical recovery protocol over ten history-dependent RoboMME tasks. Across these tasks, \method improves micro-averaged recovery success from 7.7% to 23.4%.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Failure Recovery, Hierarchical Policies, Long-Horizon Manipulation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_LNsNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PolicyGround addresses recovery from high-level failures in frozen hierarchical vision-language-action systems. It separates identifying the remaining task objective from constructing an instruction that the particular target policy can execute. Its central mechanism, Chain-of-Probe, alternates high-level subtask prediction and low-level action queries, using observations retrieved from the failed episode as proxy contexts for subsequent queries. A frozen multimodal supervisor evaluates these responses and revises the recovery instruction before a single recovery execution is committed.

The paper introduces RoboMME-FR, a recovery protocol covering ten history-dependent simulation tasks. Across 273 failures from 500 nominal episodes, PolicyGround increases recovery success from 7.7% to 23.4% and end-to-end success from 49.6% to 58.2% relative to direct semantic recovery. The evaluation includes feedback-channel ablations, revision-depth analysis, latency measurements, and a human-instruction reference.

Critical strengths
  1. A useful recovery interface with a concrete mechanism for obtaining policy-specific feedback.

The paper targets an important deployment constraint: improving recovery without retraining the target hierarchy, accessing its internal representations, or physically testing multiple candidate instructions. Reusing failed-episode contexts to continue black-box policy queries is a distinctive design within this setting. Figure 5 provides an informative example: an instruction names the correct target but produces ineffective low-level behavior, while feedback-driven insertion of a release prerequisite enables recovery. This demonstrates why inspecting the target policy can be useful beyond identifying a semantically correct goal. The authors also explicitly acknowledge that the resulting proxy response is neither a dynamics model nor a physical-feasibility guarantee.

  1. A meaningful empirical improvement under a reasonably controlled recovery protocol.

The principal comparison uses identical saved failure states, the same context-reset interface, and the same recovery horizon. Table 1 corresponds to 21 successful recoveries for semantic recovery and 64 for PolicyGround: an additional 43 successes, with improvements on eight tasks and ties on two. This is a substantive result. The denominator includes all 273 failures, including five episodes without an initial candidate. The human reference is constructed before execution and is appropriately described as an empirical reference rather than a theoretical upper bound. The paper also reports unfavorable findings, including the stronger recovery rate of low-level-only feedback and the small benefit of additional revision.

Critical weaknesses
  1. The experiments do not isolate the contribution of the central multi-step probing mechanism.

The automatic baseline directly executes the initial semantic instruction, whereas PolicyGround adds policy queries, supervisor evaluations, instruction revision, and candidate selection. Consequently, the main result establishes the effectiveness of the complete pipeline but does not determine which additions are necessary.

The feedback ablations in Tables 2 and 8 do not resolve this: they change the feedback supplied to the reviser and the selection score, rather than directly removing proxy-context chaining. The paper lacks both a budget-matched semantic self-refinement baseline without policy feedback and a first-step or fixed-context probing baseline. These are particularly consequential controls because the claimed methodological advance is obtaining useful multi-step evidence before execution.

Table 10 shows that one policy-guided revision already achieves 22.7% recovery, compared with 23.4% for full search—only two additional successful episodes. This does not establish that one-step probing is sufficient, since revision depth and probe depth are different quantities. It does, however, leave the value of the more elaborate search and chaining components unresolved. Establishing a clear advantage over these simpler controls is central to assessing the contribution.

  1. The reliability of the proxy evidence and the interpretation of the mechanism are insufficiently validated.

Appendix B.2 retrieves historical observations by matching the gripper’s open/closed state and finding the nearest recorded end-effector position within 50 mm. This does not check whether the robot is holding the same object, whether objects occupy the same locations, or whether the task has reached the same stage. Consequently, a retrieved observation may match the robot’s pose while depicting a different task state. This matters because subsequent policy queries depend on that observation.

When retrieval fails, the method can still query the high-level policy using the last available context, which may not reflect completion of the preceding action. The low-level evaluator also permits favorable scores for incomplete predicted motions. For example, Figure 5 shows a revised pickup whose predicted trajectory ends before the grasp, while the aggregate low-level score is 0.95. Such feedback may identify a promising action, but it does not establish that the full action can be completed. The paper should report how often retrieval yields relevant continuations, how frequently probes end prematurely, and whether the additional responses obtained through chaining help identify candidates that succeed during actual execution.

The HA results provide only limited evidence about the source of the improvement. Appendix D.4 counts a sequence as aligned if it contains a nonempty correct prefix of the required checkpoints, even when later steps are missing. It also excludes some preparatory actions. Thus, adding a necessary put-down step could improve recovery without changing HA. The nearly unchanged HA cannot identify which aspects of instruction sequencing or execution account for the gain. Similarly, the human reference’s 95.2% HA and 30.8% recovery suggest a substantial gap between checkpoint alignment and successful completion, but do not establish how much of that gap comes from low-level execution. These concerns limit the mechanistic interpretation.

  1. The absence of real-robot experiments leaves the framework’s practical effectiveness unverified, particularly whether recovery instructions selected through historical proxy contexts remain effective under real-world perception errors, contact uncertainty, and discrepancies between predicted and executed actions.
Questions for authors

Under the same supervisor, failed-episode evidence, interface description, initial candidate, and candidate budget, how does PolicyGround compare with semantic self-refinement without policy feedback? Please report supervisor calls or token usage so that the additional reasoning budget is visible.

How does full Chain-of-Probe compare with querying only the first high-level subtask and low-level action chunk, and with probing while keeping the failure observation fixed? For the full chain, can you report retrieval/truncation statistics and an offline comparison between proxy feedback and actual candidate execution outcomes?

Among HA-aligned but unsuccessful episodes, how many contain a complete correct high-level plan, and how many contain only a correct prefix followed by omission or termination? Does the claimed downstream-execution explanation remain supported after separating these cases?

Are there any real-world experiments that validate your method beyond simulation?

1 comment
AuthorsOpenReview ↗

Response to Reviewer LNsN

Thank you for the detailed and constructive review, and for recognizing both the recovery interface and the controlled evaluation protocol. Your summary of the main result (21 -> 64 successful recoveries) is correct.

  • W1 (isolating multi-step probing). We agree. Our experiments establish the effectiveness of the full pipeline but do not isolate proxy-context chaining. As you note, revision depth and probe depth are different quantities, so Table 10 does not show that chaining is necessary. A budget-matched self-refinement baseline and first-step / fixed-context probing are the right controls, and we will add them in a future revision.

  • W2 (reliability of proxy evidence). We agree that pose-and-gripper retrieval can return a context depicting a different task state. Sec. 3.3 states that the proxy response is neither a dynamics estimate nor a feasibility guarantee. The low-level evaluator scores truncated chains on their observed prefix only, so that the hop budget is not penalized. We agree this can be optimistic, as in Fig. 5. On HA, we agree with your characterization: prefix-based alignment cannot attribute the gain to downstream execution. We will soften the statement that the gain "primarily reflects improved compatibility with downstream execution."

  • W3 (real robot). We acknowledge the absence of real-robot experiments and will state this limitation explicitly.

Reviewer_2whtOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how to recover from high-level failures in the hierarchical VLA setting. The paper aims to recover from high-level failures by issuing semantic corrections. To ensure the effectiveness of the corrections, the paper proposes using a chain-of-probe method, where it tries to test the effectiveness of the correction on past experiences during execution and modify the correction if needed. The paper tests its proposed method on the RoboMME-FR benchmark and shows a significant recovery rate compared to baselines.

Critical strengths
  1. The paper proposes a useful problem setting, a clearly specified inference pipeline, and a substantial recovery improvement over its chosen baseline.
Critical weaknesses
  1. There are no real-robot experiments, which I think is critical for a robotics-related paper.

  2. There is no ablation study on the number of steps for chain-of-probe. E.g., is a chain really needed, or is just one step enough?

  3. The retrieval of the relevant states/images from the execution history seems to be too idealistic. In real cases, the execution history may not cover the most relevant retrieval target. Moreover, the retrieval is based only on the robot state, but the environment state may have changed a lot.

  4. The proposed method incurs significant execution time overhead.

  5. While being training-free and general, the paper is currently restricted to the VLA setting. The proposed method may be more suitable for the WAM setting, as imagined states can be obtained.

  6. The evaluation is restricted to Qwen3-VL-4B, as it sticks with the previous setup on RoboMME. A related weakness is that it is evaluated only on RoboMME.

Questions for authors

Please address the weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 2wht

Thank you for the critical and concrete feedback, and for recognizing the problem setting, the clearly specified pipeline, and the recovery improvement.

  • W1. (real robot). We acknowledge that we have no real-robot results. We will state this scope limitation explicitly and calibrate our claims accordingly.

  • W2. (chain length). We agree that an ablation on probe depth is missing. Table 10 varies revision depth, not probe depth, so it does not settle this question. We will add this ablation in a future revision.

  • W3. (retrieval). We agree that the failed episode may not cover the relevant states, and that matching on robot state does not check the environment state. When no recorded context is close enough, the chain stops rather than fabricating a state. As noted in Sec. 3.3, the resulting response is not a dynamics estimate.

  • W4. (overhead). Generation takes 101 s, compared with 45 s for semantic recovery (Table 1). This cost is incurred once per failure, before execution, and uses only offline policy queries. A single revision already reaches 22.7% RS, versus 23.4% for full search (Table 10), so a smaller candidate budget is a cheaper operating point.

  • W5. (WAM setting). Yes, we agree this is a natural extension. Chain-of-Probe needs only a source of next observations. Retrieval is the training-free instance of that source, and an action-conditioned world model could replace it without changing the rest of the pipeline. We will add this discussion.

  • W6. (Single policy and benchmark). We additionally evaluated the MemER-based hierarchical policy released with RoboMME, using the same supervisor, prompts, interface descriptions, and budgets. Recovery success (RS) and end-to-end task success (E2E) over 253 nominal failures:

MethodRSE2EGen. time
Semantic recovery8.3%53.6%42.4 s
PolicyGround11.5%55.2%127.6 s
Human instruction (ref.)20.9%60.0%–

PolicyGround improves RS on 5 tasks, ties on 4, and loses one episode on VR. The gain is smaller than with GroundSG+QwenVL. We read this as evidence that the procedure transfers without re-engineering, not that the gain is policy-independent. We chose RoboMME because it provides history-dependent tasks together with released hierarchical policies whose levels can be queried separately. We acknowledge that evaluating on a single benchmark is a limitation.

Reviewer_W1woOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents POLICYGROUND, a training-free framework for recovering from high-level failures in frozen hierarchical VLA systems. Its Chain-of-Probe mechanism queries the target policies and retrieves observations from the failed episode to construct proxy multi-step responses. An external supervisor uses these responses to evaluate and revise recovery instructions before execution. Results on ten RoboMME-FR tasks show improved recovery over directly executing a semantic recovery instruction.

Critical strengths

The distinction between semantic correctness and policy-supported recovery is useful. A recovery instruction may identify the correct goal while omitting prerequisites or using expressions that the target policy cannot reliably execute. The paper addresses this gap explicitly.

Critical weaknesses
  1. Retrieval is restricted to the current failed episode, while recovery may require states absent from that history. Moreover, matching end-effector position and gripper state does not ensure consistent object configurations or grasp relationships.
  2. The method relies on several conditions that may limit its applicability: the failed episode must contain useful proxy observations for recovery, the target system must expose separately queryable high- and low-level policies, and the supervisor receives task-specific instruction templates and interface descriptions. Failure detection is also provided externally. These conditions may be difficult to satisfy when recovery requires previously unseen states or skills unsupported by the frozen policy.
Questions for authors
  1. What evidence supports the assumption that a single failed episode provides sufficient proxy contexts for recovery? Please report retrieval failure and probe truncation rates, and clarify how the method handles recovery steps whose resulting states are absent from the recorded history.
  2. Matching end-effector position and gripper state does not guarantee consistency in object configurations or grasp relationships. How reliably do compatibility judgments based on retrieved contexts predict actual execution outcomes, particularly when these aspects differ?
  3. To what extent do the gains depend on task-specific instruction templates and detailed interface descriptions? What evidence suggests that the approach would transfer to a different hierarchical policy without substantial adaptation of these specifications?
1 comment
AuthorsOpenReview ↗

Response to Reviewer W1wo

Thank you for the review, and for recognizing the distinction between semantic correctness and policy-supported recovery as useful.

  • Q1/W1 (proxy contexts from a single episode). We agree this is a core limitation. We use the failed episode because it is the only training-free source of in-distribution observations of the same scene. When no recorded context lies within the retrieval threshold, the chain stops rather than fabricating a state (Alg. 2). The evaluator then scores only the observed prefix, so such candidates receive less evidence, not invented evidence. Recovery steps whose resulting states are absent from the history are therefore probed only partially.

  • Q2 (EEF/gripper matching vs. object configuration). We agree. As stated in Sec. 3.3, the retrieval-stitched response is not an estimate of environment dynamics, and its scores are heuristic compatibility signals rather than success predictors. We will make this limitation more prominent and add examples where the retrieved object configuration differs from the actual one.

  • Q3/W2 (templates, interface, applicability). The interface description DπD_\pi comes from the original benchmark task definitions and contains no episode-specific information. The semantic-recovery baseline receives the same DπD_\pi, so the reported gain does not come from DπD_\pi itself. On a second hierarchical policy (MemER-based), PolicyGround with the same DπD_\pi still improved recovery, though by a smaller margin; please see our response to Reviewer 2wht. Separately queryable high-/low-level policies are a scope assumption, and external failure detection is complementary (Sec. 2). We will state both explicitly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.