- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
The paper proposes a policy-aligned predictive verifier (PAV) that predicts the consequences of candidate action prefixes in a VLA policy’s internal representation space and estimates their compatibility with successful task continuation. These estimates determine how many actions to execute before obtaining a new observation and replanning. The verifier is trained using trajectory-level success/failure supervision, with a localization prior intended to preserve valid behavior within failed rollouts. Evaluations with Pi0 and Pi0.5 on LIBERO, RoboTwin 2.0, and real-world manipulation tasks show improved trade-offs between task success and policy-query cost.
Critical strengths
- The paper asks whether executing a proposed prefix preserves successful continuation, rather than relying solely on action-generation stability or prediction agreement. This directly connects horizon selection to the task objective.
- Learning local judgments from coarse supervision. The trajectory-level objective and localization prior provide a way to train the verifier from terminal outcomes without requiring action-level validity annotations.
- Experiments span two VLA policies, simulation benchmarks, and real-world tasks. Ablations examine future prediction, policy representations, loss terms, and score normalization. The reported additional GPU computation is small.
Critical weaknesses
- The improvement over the strongest adaptive baseline is modest. PAV achieves slightly higher success than AutoHorizon in the reported LIBERO H=50 settings but requires more policy queries. The advantage at comparable query budgets remains unclear.
- Performance shows a sharp peak at the default acceptance threshold, raising concerns about robustness to this hyperparameter.
- Several real-world implementation and evaluation details remain unspecified, limiting reproducibility.
Questions for authors
- How was beta=0.90 selected for the threshold, and was this choice made exclusively using validation data? how robust is its performance across repeated runs given the sharp peak at beta=0.90?
- Have you directly tested whether low-scoring action extensions reduce subsequent task success, or is the verifier validated only through overall execution performance?
- Does inference and verification block robot execution or overlap with ongoing motion? Please clarify the real-world details marked “TBD” and report end-to-end latency if available.
7 comments
[Q2] Direct Evaluation of Low-Scoring Action Extensions
Thank you for this helpful question. We directly evaluated whether executing low-scoring action extensions reduces subsequent task success rates, separately from overall execution performance. We focused on this analysis and conducted comparisons across candidates and execution lengths as supporting analyses.
We randomly sampled states within each suite from episodes used for neither training nor model selection. We then selected states based solely on variation in continuation success rates across candidates or prefix lengths. This selection did not use verifier scores or their agreement with continuation outcomes. This selection did not use verifier scores or their agreement with continuation outcomes. At each state, we computed scores for eight action candidates before execution. We then restored the same simulator state to execute each prefix and performed eight continuations per candidate–length pair using a common FixedH10 continuation policy. Repetition seeds, success criteria, and episode time limits were identical across comparison conditions. Thus, the evaluation targets were newly measured continuation success rates after each prefix, rather than terminal labels from existing trajectories.
Our primary analysis compared the local score of each additional five-action segment with the corresponding change in success rate. For the same starting state and candidate, we computed the difference between the success rates after executing prefixes of length and , given by . Holding the starting state and extension interval fixed, we calculated the Spearman correlation between the segment’s local score and across eight candidates. Across 100 interval comparisons, the median correlation was 0.3853. This supports an association between lower local scores and less favorable changes in success rate.
For example, for an H15 H20 extension in the Goal suite, a low-scoring candidate and the highest-scoring candidate both had a success rate of 4/8 before extension. After extension, the low-scoring candidate’s success rate decreased to 1/8, while the highest-scoring candidate’s remained at 4/8. The correlation across all eight candidates in this comparison was 0.9262. Similarly, for an H20 H25 extension in the Object suite, a low-scoring candidate’s success rate decreased from 8/8 to 6/8, whereas the highest-scoring candidate’s increased from 3/8 to 6/8. The correlation across all eight candidates was 0.8225.
We also conducted two supporting analyses of cumulative prefix scores.
- Comparing candidates at a fixed execution length, we held the starting state and prefix length fixed to assess whether the score distinguishes candidates with different continuation success rates. Across 100 comparisons, the Spearman correlation between prefix scores and continuation success rates had a mean of 0.3383 and a median of 0.2943.
- Comparing execution lengths for a fixed candidate, we held the starting state and candidate fixed and compared H10, H20, and H30. Across 100 comparisons, the mean correlation was 0.7386, with a median of 0.8660.
These experiments provide direct validation beyond overall system performance. The supporting analyses show that prefix scores track subsequent success rates, while the extension analysis directly demonstrates an association between lower local scores and less favorable changes in success rate. Although evaluated using a FixedH10 continuation policy, these results provide evidence that the learned scores reflect meaningful variation in actual continuation success rates, including changes associated with executing additional action segments. We will clarify these findings and the supporting evidence in the paper.
[W1] Success–Query Trade-offs and Generalization Across Benchmarks
We agree that PAV sometimes requires more queries than AutoHorizon and that the difference in success rate can be modest. As discussed in our response to Q1, we used a fixed threshold across environments to demonstrate that a shared threshold can achieve strong performance. Using suite-specific or task-specific thresholds achieves slightly higher success rates with fewer queries than AutoHorizon. In the same LIBERO––H50 setting, AutoHorizon achieves a 94.45% success rate with 12.7 queries on average.
We also examined whether AutoHorizon’s reported RoboTwin performance extends to a shared policy checkpoint across tasks. The AutoHorizon paper reports results on seven RoboTwin tasks using checkpoints fine-tuned separately for each task. These results therefore do not directly establish its performance with a shared checkpoint trained across RoboTwin tasks. For a fair comparison, we evaluated AutoHorizon using the same checkpoint as PAV under matched evaluation conditions. The results are shown below.
\begin{array}{lrrr} \text{Method} & \text{Success rate}\uparrow & \text{Average queries}\downarrow & \text{Average execution horizon} \\ \hline \text{PAV} & 58.72\% & 56.75 & 10.25 \\ \text{AutoHorizon (reimplementation)} & 45.64\% & 14.53 & 29.77 \end{array}
In this setting, AutoHorizon predominantly selects the maximum horizon of 32, yielding an average execution horizon of 29.77. This suggests limited effective adaptation under the shared-checkpoint setting. Its success rate of 45.64% is also close to the 45.40% achieved by fixed H32 execution. In contrast, PAV uses a shorter average execution horizon of 10.25 and achieves a success rate 13.08 percentage points higher, while requiring more policy queries.
Together, these results support PAV’s effectiveness across multiple benchmarks. On LIBERO, suite-specific or task-specific thresholds improve success rates while reducing queries relative to AutoHorizon. On RoboTwin, PAV achieves a larger success-rate improvement at the cost of more queries.
[W3,Q3] Real-World Execution and Evaluation Protocol
Thank you for raising this important point. Our real-world system uses a sequential execution scheme without overlap between action execution and inference. After completing the action prefix selected at the previous decision, the system obtains a new observation, performs policy inference and verification, and then starts executing the next prefix. Execution of the next action therefore waits until inference and verification are complete. We apologize for not being able to report end-to-end latency, as we did not separately record this measurement.
We will also fill in the TBD entries in Appendix E.1–E.2. Success is judged by a human evaluator. For Tray, success requires releasing the cube from the gripper and placing it correctly inside the tray. For Stack, the cubes must remain stacked without collapsing after the gripper releases the cube. For Button, the button must be pressed and produce an audible click. The maximum execution budget is 500 action steps per trial, with no separate wall-clock time limit.
For each task, we use the checkpoint obtained after 10,000 updates of full fine-tuning of , and train each verifier for 60 epochs on 50 rollouts. The candidate horizons are , the shared threshold is , and square-root block-count normalization is applied. At each decision, we generate a single action chunk and execute the selected prefix without regenerating the chunk. Evaluation consists of 30 trials per task, separate from the training data, with initial object configurations matched across methods by episode index. Query counts include the initial generation, and the average horizon is computed by assigning equal weight to every selected prefix.
[W2,Q1] Threshold Selection and Sensitivity
We apologize for not clearly explaining the selection of β=0.90 and the sensitivity to this threshold. We selected the threshold using only validation data separate from the final evaluation data. Specifically, for each sample, we used the minimum of its first two prefix scores as its representative score. Because the first horizon extension is a shared prerequisite for all subsequent extensions, this choice aims to calibrate the criterion for extending execution beyond the minimum length. We then applied Kittler’s method, which approximates the score distribution with two Gaussians and minimizes the separation error. The resulting thresholds were 0.906771 for LIBERO–π0–H50, 0.900278 for LIBERO–π0.5–H10, 0.898986 for LIBERO–π0.5–H50, and 0.907220 for RoboTwin–π0.5–H32. To evaluate transferability across tasks and environments, we applied a common threshold of β=0.90 in all experiments. These results support the applicability of a shared threshold in the evaluated environments, but do not establish its optimality in every environment. Also, to address your question about robustness across repeated runs, we evaluated PAV at β=0.90 across three evaluation seeds. Its mean success rate was 94.88% (± 0.45%). This provides evidence of stable performance at the default threshold across these runs.
We additionally compared the unrounded shared, suite-specific, and task-specific thresholds on LIBERO–π0.5–H50, as shown below. The suite-specific thresholds were 0.9123 for Spatial, 0.8961 for Object, 0.9147 for Goal, and 0.8892 for Long. Nevertheless, we acknowledge that threshold sensitivity is a limitation of our method.
\begin{array}{lrrrrrr} \text{Threshold setting} & \text{Spatial} & \text{Object} & \text{Goal} & \text{Long} & \text{Overall success rate} & \text{Average queries} \\ \hline \text{Shared Kittler }(\approx 0.899) & 97.0\% & 98.6\% & 91.8\% & 91.2\% & 94.65\% & 13.74 \\ \text{Suite-specific Kittler} & 96.8\% & 99.0\% & 91.4\% & 92.2\% & 94.85\% & 11.92 \\ \text{Task-specific Kittler} & 96.4\% & 98.4\% & 94.6\% & 93.0\% & 95.60\% & 12.38 \end{array}
The substantial performance variation around β=0.90 can be understood through the trade-off between the temporal consistency of longer execution and the responsiveness of shorter execution. A lower threshold permits longer execution, but may delay necessary reactions, allowing errors to accumulate and leading to states from which success is difficult. We also observed more frequent short executions following long executions, consistent with attempts to recover from accumulated errors. Specifically, lowering β from 0.90 to 0.88 increased the proportion of h≥20 selections from 21.88% to 42.92% and that of h=5 selections from 24.99% to 36.59%, while reducing the success rate from 94.65% to 82.75%. Furthermore, the number of episodes in which a long execution (h≥20) was followed by low scores for h=5 for at least three consecutive observation steps increased from 59 at the reference threshold (β≈0.899) to 97 at β=0.88. This is consistent with repeated short executions in states assessed as having a low likelihood of success after a long execution. It suggests that even frequent h=5 selection may not prevent performance degradation if the system fails to respond when needed.
Conversely, a higher threshold imposes stricter conditions on execution extension. Although shorter execution provides responsiveness, it limits the benefits of temporal consistency and query reduction offered by longer execution. The method still selects some longer action chunks and outperforms fixed h=5 execution, but appears unable to fully exploit these benefits. Specifically, at β=0.92, the h=5 selection rate was 53.12%, and the average number of queries increased from 14.0 at β=0.90 to 25.1. The success rate decreased from 94.65% to 88.30%, although it remained above the 87.15% achieved with fixed h=5 execution. These results are consistent with the trade-off discussed in the introduction. They highlight the importance of choosing a threshold that provides longer execution for consistency and shorter execution for responsiveness when each is needed.