PaperlayerView exact revision on arXiv

Source-linked AI summary

Policy-Aligned Predictive Verification for Adaptive VLA Action-Chunk Execution

Minkyu Kim, Junhoo Lee, Seungyeon Kim, Suhyun Jeon, Baekseung Kim, Donghoon Han, Nojun Kwak

arXiv:2610.21434v1cs.ROcs.LG

TL;DR

Adaptive execution needs to determine whether a proposed action prefix preserves the possibility of completing the task, which generation stability and prediction agreement alone do not establish. PAV predicts prefix consequences in the VLA’s task-conditioned representation space and uses validity scores learned from terminal outcomes to select execution horizons. Across simulated and real-world manipulation, results support learning execution decisions from task outcomes rather than consistency alone.

  • Problem

    Adaptive execution must choose prefixes that preserve successful task completion, but generation stability and prediction agreement alone do not establish that compatibility.

    Source: §1 · pp. 1–2

  • Method

    PAV predicts candidate-prefix consequences in the VLA’s task-conditioned representation space and uses outcome-supervised validity scores to select execution horizons.

    Source: §1 · p. 2, §3.3 · p. 5

  • Results

    Predicting future VLA representations achieves higher success with fewer policy queries than direct verification or prediction in generic visual features across evaluations with π0 and π0.5.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Across simulated and real-world manipulation, results support learning execution decisions from task outcomes rather than relying on consistency alone.

    Source: §5 · p. 9

    Proposal validity is assessed with task progress and remaining time in the state, under a fixed continuation protocol.

    Source: §3.2 · p. 4

Abstract

from arXiv · show

Vision-language-action (VLA) policies generate action sequences that reduce repeated inference, but deciding how much of each sequence to execute is critical to task success. Existing adaptive methods make this decision using generation stability or prediction agreement, yet these signals do not establish whether executing a candidate prefix preserves the ability to complete the task. We introduce a policy-aligned predictive verifier (PAV) that anticipates the consequences of candidate action prefixes and evaluates their compatibility with successful continuation before execution. Rather than using future prediction to generate actions, our method evaluates existing proposals in the VLA's internal task-conditioned representation space. To learn local judgments from terminal task outcomes, we combine trajectory-level survival supervision with a localization prior that preserves useful behavior within failed rollouts. The resulting validity scores guide execution-horizon selection. Evaluations with pi_0 and pi_0.5 across LIBERO, RoboTwin 2.0, and real-world manipulation show improved success-query trade-offs. These findings point toward a different role for future prediction in robotics, using it to assess how far to trust a policy's proposed actions rather than to generate new ones.

1 INTRODUCTION

Adaptive execution must choose a prefix that preserves successful task completion, a property that stability and prediction-agreement signals alone do not establish. PAV predicts the consequences of proposed prefixes in the VLA’s task-conditioned representation space and uses validity estimates to guide execution horizons.

1 INTRODUCTION

  • Adaptive execution seeks the longest action-chunk prefix that still allows successful task completion, but its effects on that possibility are not directly observable before execution.Generation stability and prediction agreement are indirect signals and do not establish compatibility with successful completion.Source: §1 · pp. 1–2
  • Future prediction evaluates existing proposals in the VLA’s task-conditioned representation space rather than generating new actions.The representation integrates visual observations and language instructions.Source: §1 · p. 2
  • PAV assesses each action extension’s compatibility with successful continuation from predicted VLA representations, then combines local validity scores to select execution horizons.Terminal outcomes supervise accumulated validity, while a localization prior concentrates score reductions in fewer extensions so failed rollouts can retain valid prefixes.Source: §1 · p. 2
  • Predicting future VLA representations achieves higher success with fewer policy queries than direct verification or prediction in generic visual features.Experiments with π0 and π0.5 span LIBERO, RoboTwin 2.0, and real-world tasks.Source: §1 · p. 2

2 RELATED WORK

Prior work adapts action-chunk execution using stability or prediction agreement, uses future models for planning and generation, and derives local learning signals from global outcomes. This work instead applies prefix-conditioned prediction and trajectory outcomes to verify proposals and select execution horizons.

2.1 ADAPTIVE EXECUTION OF ACTION CHUNKS

  • Adaptive-execution methods use generation stability or prediction agreement, whereas this verifier predicts candidate-prefix consequences and assesses compatibility with successful continuation before execution.AutoHorizon uses action self-attention; A3 uses action consensus and re-decoding stability; VLA-Corrector and CheckVLA use prediction-observation discrepancies [Wang et al., 2026; Chen et al., 2026; Pan et al., 2026; Liu et al., 2026].Source: §2.1 · p. 2

2.2 WORLD MODELS AND PREDICTIVE REPRESENTATIONS

  • Unlike world models used for planning or action generation, this approach predicts prefix-induced futures in the action-generating policy’s representation space for proposal verification.The representation retains joint observation and instruction context relevant to control.Source: §2.2 · p. 3

2.3 LEARNING FROM TRAJECTORY-LEVEL OUTCOMES

3 METHOD

The method chooses action-chunk prefixes by estimating whether they preserve compatibility with successful task continuation, rather than relying only on generation stability or prediction agreement. It predicts future policy representations, verifies local validity, and uses the resulting scores to select an execution horizon.

3.1 PRELIMINARIES

  • Generation stability and prediction agreement offer indirect execution signals, but neither establishes whether a candidate prefix preserves compatibility with successful continuation.Adaptive execution trades inference cost and temporal consistency against responsiveness when choosing when to replan.Source: §3.1 · pp. 3–4

3.2 FORMULATING PROPOSAL VALIDITY

  • Proposal validity measures whether successful continuation remains possible throughout a candidate prefix, with task progress, remaining time, and a fixed continuation protocol shaping that judgment.Prefix-closed compatibility gives validity a cumulative structure for deciding how far execution can proceed.Source: §3.2 · p. 4

3.3 LEARNING PREDICTION AND VERIFICATION

  • The predictor forecasts future features in the VLA’s task-conditioned representation space, and an action-sensitivity margin encourages executed actions to produce predictions closer to observed futures than alternative prefixes.The prediction targets are representations from executed transitions; alignment alone may let the predictor ignore which actions produced the future.Source: §3.3 · pp. 4–5
  • The verifier estimates local validity from predicted representations and accumulates those estimates into prefix scores, while terminal rollout outcomes supervise the trajectory-level product rather than directly labeling each extension.Successful rollouts raise every factor; failed rollouts lower only the product, allowing compatible extensions to retain high scores. A localization prior addresses diffuse score decreases across failed rollouts.Source: §3.3 · pp. 5–6

3.4 EXECUTION-HORIZON SELECTION

  • Execution-horizon selection length-normalizes accumulated validity, checks candidate prefixes in ascending order, and stops at the first score below β.The robot executes the selected prefix and queries the policy again from the resulting observation.Source: §3.4 · p. 6

4 EXPERIMENTS

Across LIBERO, RoboTwin 2.0, and real-world tasks, adaptive execution improves success-query trade-offs over fixed horizons. Prediction and policy-alignment analyses support using predicted task-conditioned representations to assess candidate actions.

4.1 EXPERIMENTAL SETUP

  • Evaluation covered π0 and π0.5 across LIBERO suites, all 50 RoboTwin 2.0 tasks, and three real-world tasks, with success and mean VLA queries used to assess performance against inference cost.Real-world tests used task-fine-tuned π0.5 policies for placement, stacking, and contact tasks.Source: §4.1 · p. 7, §4.3.1 · p. 9

4.2.1 LIBERO

  • Across all three LIBERO policy–chunk settings, our method achieved the highest average success; at π0.5, H = 50, it gained 7.50 percentage points over h = 5 while reducing queries by 63.3%.At H = 10, success rose from 97.50% to 98.10% with similar query counts.Source: §4.2.1 · p. 8

4.2.2 ROBOTWIN 2.0

  • On all 50 RoboTwin 2.0 tasks, our method reached 58.64% success versus 55.76% for the best fixed horizon, h = 8, with similar query counts of 58.5 versus 54.5.On the selected 16 tasks, success was 54.13%, exceeding the best fixed horizon (47.88%) and DVAC (41.6%).Source: §4.2.2 · p. 8

4.2.3 ANALYSIS OF PREDICTION AND POLICY ALIGNMENT

  • On held-out LIBERO rollouts, predictions conditioned on executed actions reached 0.9401 cosine similarity, versus 0.5304 with same-task swapped actions.Explicit prediction raised success from 84.55% to 94.65% and reduced mean queries from 22.3 to 14.0 versus Direct; policy-space prediction achieved higher success and lower query cost than DINOv2 features.Source: §4.2.3 · pp. 8–9, Table 4 · p. 8
  • Verifier GPU computation took 0.90–2.49 ms, or 0.45–1.24% of action-generation time, but these measurements do not represent end-to-end control-loop latency.The analysis compares prediction and policy-alignment variants on π0.5 with H = 50, using ten candidate execution lengths.Source: §4.2.3 · pp. 8–9, Table 5 · p. 8

4.3.2 REAL-WORLD RESULTS

  • 74.4% average real-world success versus 41.1% for the strongest fixed horizon, h = 10, while mean replanning fell from 36.54 to 12.87 per episode.At a similar replanning count, adaptive execution reached 74.4% success versus 40.0% for h = 30.Source: §4.3.2 · p. 9, Table 6 · p. 9

5 CONCLUSION

The paper frames action-chunk execution as a commitment-length decision and uses future prediction to inform that commitment in the task-conditioned representation space used for action generation. Results across simulated and real-world manipulation support learning execution decisions from task outcomes rather than consistency alone.

5 CONCLUSION

  • Future prediction evaluates existing proposals in the task-conditioned representation space that guides action generation, helping decide how long to commit to action chunks.The authors describe predictive verification as a way to anticipate proposal consequences before acting.Source: §5 · p. 9

AI USE STATEMENT

  • Generative AI tools assisted implementation after methodology development and helped draft and edit the manuscript; authors state they reviewed and verified all assisted text and code.They report that the tools were not used to develop methodology, formulate hypotheses, or interpret results.Source: AI USE STATEMENT · p. 10

A IMPLEMENTATION AND CONFIGURATION

The verifier predicts future policy representations from action chunks, then scores those predicted consequences for local validity. Its inputs preserve policy context while using a learned projection to make prediction more manageable.

A.1 ARCHITECTURE OF THE PREDICTIVE VERIFIER

  • The verifier represents policy context using pooled final-prefix language-model states and preprocessed state, concatenated into a 6,176-dimensional vector while leaving VLA parameters unchanged.The instruction enters through policy language tokens; no separate verifier task-ID embedding is used in the described configuration.Source: §A.1 · p. 13
  • A learned projector maps standardized 6,176-dimensional context to a normalized 128-dimensional policy-feature space, where the predictor and verifier operate.The current and observed future contexts share the projector, and the future target is detached in prediction losses.Source: §A.1 · p. 13
  • For π0.5 with H = 50, the predictor processes ten five-action blocks recurrently, updating the normalized latent with each action-conditioned prediction; teacher forcing is disabled.The local verifier reads only the predicted 128-dimensional future latent and outputs a sigmoid validity score.Source: §A.1 · p. 13, Table 7 · p. 13

A.2 TRAINING DATA AND OPTIMIZATION

Training uses policy rollouts with varied execution horizons, terminal outcomes, and observed future contexts to train prediction and local validity judgments. Deployment evaluates candidate horizons sequentially and stops at the first score below threshold.

A.2 TRAINING DATA AND OPTIMIZATION

  • Rollouts sampled horizons uniformly from {5, 10, . . . , 50}; the 6,000 episodes were split by initial-state identity into 4,200 training, 840 validation, and 960 test rollouts.Training samples balance successful and failed outcomes, and normalization statistics use only the training partition.Source: §A.2 · pp. 13–14
  • Future alignment trains predictions against detached observed future latents, while same-task action-swap ranking encourages sensitivity to the proposed actions.Swapped-action comparisons use a 0.1 margin, and records without another same-task record are excluded.Source: §A.2 · p. 14
  • Trajectory supervision treats successful rollouts as support for validity across extensions and failed rollouts as evidence that at least one extension is incompatible; localization preserves useful blocks within failures.An auxiliary terminal-context completion objective is weighted 0.25 and does not enter the execution score.Source: §A.2 · p. 14, Table 8 · p. 14

A.3 CANDIDATE HORIZONS AND EXECUTION

  • Candidate horizons span each chunk at regular intervals, and deployment scans them in ascending order, stopping at the first accumulated score below 0.9.The H = 50 configuration uses five-action prediction strides and square-root length normalization; intermediate checks prevent extending past an earlier violation.Source: §A.3 · pp. 14–15, Table 9 · p. 15

B DETAILED ROBOTWIN RESULTS

The RoboTwin analysis reports per-task comparisons under a distinct verifier configuration, while the broader ablation analysis uses controlled comparisons to characterize the H = 50 LIBERO setting.

B DETAILED ROBOTWIN RESULTS

  • The RoboTwin verifier uses π0.5 with H = 32, β = 0.91, and a task embedding; its h = 32 baseline is separately evaluated, so episode initial conditions are not guaranteed to be paired.DVAC is an external reference using its published H = 50 policy and evaluation protocol.Source: B DETAILED ROBOTWIN RESULTS · p. 15

C ADDITIONAL ABLATIONS AND ANALYSIS

  • The ablations focus on π0.5 with H = 50 across four LIBERO suites, offering ten candidate horizons from 5 to 50 steps; only overhead measurements span multiple policy settings.Comparisons hold the checkpoint, rollout dataset, task set, and evaluation protocol fixed.Source: C ADDITIONAL ABLATIONS AND ANALYSIS · p. 15

C.1 FUTURE PREDICTION AND REPRESENTATION ABLATIONS

The section directs readers to the complete suite-level comparison and frames the individual loss ablations as isolating training-objective contributions.

C.1 FUTURE PREDICTION AND REPRESENTATION ABLATIONS

  • Table 3 compares Direct, DINOv2, and PAV using success and VLA-query counts together, distinguishing performance from replanning frequency.Individual loss ablations further isolate the contributions of training objectives.Source: §C.1 · p. 15
  • Table 10 reports per-task success rates for all 16 RoboTwin Selected 16 tasks, with bold marking the highest reported row value, including ties.DVAC reproduces its published results; other averages are computed over 800 episodes.Source: Table 10 · p. 16

C.2 ACCEPTANCE-THRESHOLD SWEEP

The analyses examine the threshold operating point, training-loss contributions, action sensitivity, and inference-time score normalization. Across these tests, the reported comparisons connect execution choices with success, query demand, or predictive responsiveness.

C.2 ACCEPTANCE-THRESHOLD SWEEP

  • 89.45% average success with 19.7 VLA queries per episode at β = 0.91 compares with 87.45% and 19.2 queries for h = 10, a 2.00-percentage-point gain at similar query cost.The threshold sweep holds the trained verifier fixed while varying β; the default setting achieves the highest success among tested thresholds.Source: Table 11 · p. 16, §C.2 · pp. 16–17
  • The threshold distributions are normalized within each setting; h ≤10 accounts for 72.6% of selections at 0.91 and 94.7% at 0.92.At the default β = 0.90, h = 15 is the most frequent choice, accounting for 33.0% of selections.Source: §C.2 · pp. 16–17, Figure 4 · p. 17

C.3 REMOVING INDIVIDUAL LOSS TERMS

  • The full objective improves success by 8.95–11.00 percentage points over the three ablated variants while reducing mean queries from 23.99–25.60 to 14.0.The comparison evaluates each variant with the default threshold and supports retaining future alignment, action-swap ranking, and localization.Source: Table 12 · p. 17, §C.3 · p. 17

C.4 SENSITIVITY TO ACTION SWAPS

  • Across horizons, executed prefixes predict observed consequences more closely than alternative prefixes in 89.3–95.2% of held-out pairs.Both inputs are evaluated against the same recorded future, making the comparison a test of action sensitivity relative to that target.Source: Table 13 · p. 18, §C.4 · p. 18

C.5 EFFECT OF SQUARE-ROOT LENGTH NORMALIZATION

  • At the shared default threshold, square-root normalization achieves the highest task success and shifts selected horizons toward intermediate lengths, with a mode at h = 15.Raw accumulation selects only h = 5 or h = 10, while geometric-mean normalization concentrates choices at h = 5 and h = 50; the normalized score is used for selection, not as a calibrated probability.Source: §C.5 · pp. 18–19, Table 14 · p. 19, Figure 5 · p. 19

D QUALITATIVE DEMONSTRATIONS

Two LIBERO demonstrations show execution horizons changing within grasp-recovery rollouts. Shorter commitments accompany retries, while longer ones appear during approach and transfer.

D QUALITATIVE DEMONSTRATIONS

  • In two LIBERO rollouts, short horizons accompany missed-grasp retries and longer horizons occur during approach and transfer.The traces mark selected horizons and policy queries; shaded intervals denote annotated interaction phases, not validity labels or training supervision.Source: D QUALITATIVE DEMONSTRATIONS · p. 19, Figure 6 · p. 20

E.1 HARDWARE, VIEWS, AND TASK LAYOUT

The real-world evaluation uses task-specific π0.5 policies and verifiers across three manipulation tasks. Qualitative examples show adaptive execution recovering from interaction errors, unlike fixed short or long horizons.

E.1 HARDWARE, VIEWS, AND TASK LAYOUT

  • The setup uses a 7-DoF Franka Panda with three RGB views, 15 Hz action execution, and a 500-step maximum per trial; human evaluators judge task success.Trials ending early without success count as failures in the success-rate denominator.Source: §E.1 · p. 21
  • Each task uses a separately fine-tuned π0.5 policy and verifier, with 50 verifier-training rollouts, 60 epochs, candidate horizons from 5 to 50, and threshold β = 0.90.Evaluation uses 30 trials per task with episode-matched object placements; each decision evaluates one chunk and executes its selected prefix without regenerating the proposal.Source: §E.1 · p. 21

E.2 REAL-WORLD EXECUTION EXAMPLES

  • Adaptive execution combines corrective feedback near interactions with sustained progress afterward, while short horizons can interrupt progress and long horizons can continue after unsuccessful interactions.The examples illustrate behavior rather than measured traces of individual policy-query events.Source: §E.2 · p. 21
  • Replanning counts include the initial chunk generation, while mean execution horizon pools selected prefixes so each selection has equal weight.Episode-level means are not averaged.Source: §E.2 · p. 21
  • PAV recovers and presses the button, realigns and stacks cubes, and returns to retry placing a cube; fixed horizons show stalls, off-target behavior, or unsuccessful continuation.The qualitative comparisons contrast h = 5 and h = 50 with PAV across all three tasks.Source: §E.2 · p. 21, Figure 7 · p. 22

Related papers

Loading 2610.21434v1…