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

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

cs.ROOct 15, 2026

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

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

Organizations: Seoul National University · KAIST · Dnotitia

Abstract

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.

Figures & tables

Figure 1: Generation stability (a) and prediction agreement (b) can remain high despite a missed grasp and task failure. Proposal validity (c) guides earlier replanning, leading to success in this example.

Figure 2: Policy-aligned predictive verification. The predictor uses the VLA representation and proposed actions to predict future policy representations. The verifier estimates local validity, which is accumulated and normalized to select an execution horizon. The robot executes the selected prefix and replans.

$\pi_{0}$ ( $H=50$ )
Method	Spatial $\uparrow$	Object $\uparrow$	Goal $\uparrow$	Long $\uparrow$	Avg. $\uparrow$	Queries $\downarrow$
$h=5$ (Baseline)	97.4	97.2	91.0	82.8	92.10	34.6
$h=10$	96.0	98.4	93.6	84.2	93.05	17.2
$h=30$	89.4	94.4	90.2	86.6	90.15	6.3
$h=50$	76.6	75.4	83.4	68.8	76.05	4.4
AutoHorizon	95.4	98.4	94.8	87.6	94.05	10.6

Table 1: LIBERO success rates (%) and mean VLA queries per episode. Our method and fixed-horizon baselines use 50 trials per task. † denotes published reference results with potentially different protocols. Bold and underlining indicate the highest and second-highest success rates within each policy setting, including ties. Dashes indicate unreported results.

Table 4

Figure 3: Experimental settings. (a) Successful real-world task states for Press Button, Stack Cubes, and Place Cube in Tray, shown from left to right. (b) Representative simulation environments from LIBERO and RoboTwin 2.0, covering single-arm tabletop and dual-arm manipulation.

Table 6

Method	Place Cube in Tray $\uparrow$	Stack Cubes $\uparrow$	Press Button $\uparrow$	Avg. success $\uparrow$	Avg. replans $\downarrow$	Avg. horizon
$h=5$	16.7	10.0	23.3	16.7	69.58	5.00
$h=10$	66.7	16.7	40.0	41.1	36.54	10.00
$h=30$	23.3	43.3	53.3	40.0	11.97	30.00
$h=50$	26.7	33.3	50.0	36.7	6.91	50.00
PAV (Ours)	80.0	66.7	76.7	74.4	12.87	30.22

Table 6: Real-world success rates (%), average replanning counts, and average execution horizons. Avg. success is the mean across the three tasks. Bold and underlining indicate the highest and second-highest success rates, including ties.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Module	Layers	Parameters
Context projector	LN, $6176\!\to\!256$ , GELU, $256\!\to\!128$ , $\ell_{2}$ normalization	1,626,560
Action encoder	LN, $35\!\to\!128$ , GELU, $128\!\to\!128$ , GELU	21,190
Recurrent transition	GRUCell, input 256, hidden 192	259,200
Residual update	LN, $192\!\to\!192$ , GELU, $192\!\to\!128$	62,144
Local verifier	LN, $128\!\to\!192$ , GELU, $192\!\to\!1$ , sigmoid	25,217
Completion head	LN, $128\!\to\!128$ , GELU, $128\!\to\!1$	16,897

Appendix

Table 7: Predictive-verifier architecture for

\pi_{0.5}

on LIBERO with

H=50

. Widths denote linear-layer dimensions. LN denotes LayerNorm. Dropout is zero throughout.

Setting	Value
Collection and training seed	7
Rollouts per task / total	150 / 6,000
Train / validation / test rollouts	4,200 / 840 / 960
Batch size	8 episodes, including their decision records
Optimizer	AdamW, constant learning rate $3\times 10^{-4}$
Training epochs / selected epoch	100 / 7

Appendix

Table 8: Training settings for the no-task-embedding

\pi_{0.5}

LIBERO

H=50

verifier.

Setting	Value
Generated action chunk	$50\times 7$
Prediction stride	5 control steps
Candidate horizons	$5,10,15,20,25,30,35,40,45,50$
Acceptance threshold	$\beta=0.9$
Length normalization	Square root of the block count
Candidate check	Stop at the first threshold violation

Appendix

Table 9: Deployment settings for the

H=50

LIBERO configuration.

Task	$h=4$	$h=8$	$h=20$	$h=32$	DVAC †	PAV (Ours)
Blocks Ranking RGB	0	40	62	38	45	64
Blocks Ranking Size	2	34	26	14	37	50
Hanging Mug	0	6	16	6	1	8
Move Stapler Pad	14	30	22	14	11	42
Open Laptop	44	62	84	60	90	70
Open Microwave	4	26	46	58	67	42

Appendix

Table 10: Per-task success rates (%) on all 16 RoboTwin tasks in the Selected 16 subset. DVAC † reproduces the

\alpha=2

results from its Table 9 ( Feng et al., 2026 ) , including the reported average rounded to one decimal place. Other averages are computed over 800 episodes. Bold indicates the highest reported value in each row, including ties.

$\beta$	Spatial	Object	Goal	Long	Avg.	Queries $\downarrow$	Mean $h$
	Success (%) $\uparrow$
0.88	79.4	87.2	84.6	79.8	82.75	13.8	15.200
0.89	81.8	90.0	85.4	85.4	85.65	15.1	13.443
Default (0.90)	96.8	98.8	91.8	91.2	94.65	14.0	12.783
0.91	89.4	94.6	87.6	86.2	89.45	19.7	9.685
0.92	90.4	94.8	86.8	81.2	88.30	25.1	7.614

Appendix

Table 11: Threshold sweep with a fixed trained verifier on

\pi_{0.5}

with

H=50

on LIBERO. Success rates are reported per suite and averaged equally across the four suites. Queries are the total histogram selection count divided by 2,000 evaluation episodes; mean horizon weights each selection equally. The default row uses the PAV results from the main paper.

Figure 4: Selected-horizon distributions for

\pi_{0.5}

with

H=50

on LIBERO. Each panel shows one threshold, with counts pooled across the four suites and normalized within that setting. Panels share the same axes, with one bar per candidate horizon. The default setting is highlighted in dark blue.

Training variant	Spatial	Object	Goal	Long	Avg.	Queries
Full objective	96.8	98.8	91.8	91.2	94.65	14.0
Without future alignment	83.2	91.0	86.8	80.2	85.30	23.99
Without swap ranking	84.6	93.4	88.2	76.6	85.70	24.26
Without localization	81.8	92.2	86.8	73.8	83.65	25.60

Appendix

Table 12: Loss-term ablations on

\pi_{0.5}

with

H=50

on LIBERO. Suite columns and the average report success (%). Queries are the mean number of VLA generations per episode across the four suites.

Horizon	$n$	Mean $\Delta_{t}(h)$	$\Delta_{t}(h)>0$ (%)	$\Delta_{t}(h)\geq m$ (%)
5	7,276	0.1628	89.3	53.5
10	6,367	0.3292	92.5	70.7
15	5,537	0.4257	93.6	76.2
20	4,683	0.4898	94.6	79.6
25	3,883	0.5287	94.7	81.0
30	3,160	0.5537	95.2	83.0

Appendix

Table 13: Horizon-wise action-swap sensitivity of the full predictor on held-out transitions.

n

is the number of observed-block pairs, and the ranking margin is

m=0.1

Normalization	Spatial	Object	Goal	Long	Avg.	Queries $\downarrow$	Mean $h$
	Success (%) $\uparrow$
Raw product ( $\eta=0$ )	92.0	90.8	91.6	80.0	88.6	30.9	6.104
Square root ( $\eta=\tfrac{1}{2}$ )	96.8	98.8	91.8	91.2	94.65	14.0	12.783
Geometric mean ( $\eta=1$ )	53.6	53.0	71.8	50.4	57.2	11.3	22.645

Appendix

Table 14: Inference-time normalization ablation with the trained predictor and verifier held fixed. Success rates are reported for each LIBERO suite and averaged equally across the four suites. Queries are the total histogram selection count divided by 2,000 evaluation episodes; mean horizon weights each selection equally. Entries correspond to the common default threshold.

Figure 5: Horizon selection under three normalization rules at

\beta=0.90

\pi_{0.5}

, LIBERO

H=50

. Each panel uses identical axes and independently normalized counts pooled across suites. Square-root normalization favors intermediate horizons. Success and query costs appear in Table 14 .

Figure 6: Adaptive execution during grasp recovery in two LIBERO rollouts. (a) In LIBERO-Object, the robot retries an empty grasp before transferring the soup can to the basket. (b) In LIBERO-Goal, it reattempts the bowl grasp and places the bowl on the stove. Observed frames are sampled every 19 and 15 control steps, respectively. The traces show selected execution horizons, with dots marking policy queries. Short horizons accompany grasp retries, while longer horizons occur during approach and transfer. Shaded intervals identify the annotated interaction phases, not validity labels or training supervision.

Figure 7: Real-world qualitative comparisons. (a) Press Button: PAV recovers and presses. $h=5$ oscillates, while $h=50$ retains an off-target alignment. (b) Stack Cubes: PAV realigns and stacks. $h=5$ grasps but stalls before transfer, while $h=50$ continues after a missed grasp. (c) Place Cube in Tray: PAV returns and retries successfully. $h=5$ drifts away, while $h=50$ continues empty before returning later without success. Red annotations highlight key behaviors. All rows start at $t=0$ . Video timestamps are nonuniform and not synchronized across rows.

Robotic Policy Evaluation

Robotic policy evaluation measures whether learned controllers succeed reliably, generalize across conditions, recover from failures, and behave safely.

See all 126 papers

PAPO-VLA: Planning-Aware Policy Optimization for Vision-Language-Action Models
May 19, 2026Peizheng Guo, Jingyao Wang, Changwen Zheng +1Robot Policy Learning Credit Assignment in RL
Institute of Software Chinese Academy of Sciences · University of Chinese Academy of Sciences
PACT-VLA: Speculative Action-Chunk Verification with Pre-Commit Observations
Oct 20, 2026Yixiang Zhu, Ruilong Wu, Yonghao Chen +3Vision-Language-Action Models Robustness of VLA Models
The Hong Kong University of Science and Technology (Guangzhou) · One Robotics
CheckVLA: Execution-Time Verification with Action-Conditioned World Model for Long-Horizon Mobile Manipulation
Jul 29, 2026Yushan Liu, Peibo Sun, Xintao Chao +8Long-Horizon Robotic Manipulation World Models for Robotics
Tsinghua University · Shanghai Jiao Tong University · Peking University +2

May 19, 2026J/K move · Enter open · S save

PAPO-VLA: Planning-Aware Policy Optimization for Vision-Language-Action Models

Peizheng Guo, Jingyao Wang, Changwen Zheng, Wenwen Qiang

Institute of Software Chinese Academy of Sciences · University of Chinese Academy of Sciences

Robot Policy Learning Credit Assignment in RL Vision-Language-Action Models RL for VLMs Closed-Loop Control of VLA Models

Open paper AskPDF arXiv

Vision-Language-Action (VLA) models show promising ability in language-guided robotic tasks. However, making VLA policies reliable remains challenging, because a manipulation task is completed through closed-loop interaction, where each action affects subsequent execution. To analyze this problem, we revisit VLA policy during execution and argue that a VLA policy acts both as a planner, which makes task-oriented decisions that change the direction of execution, and as an executor, which realizes these decisions through dense continuous actions. This view suggests that improving VLA reliability requires particular attention to planning actions. Existing optimization methods can imitate actions or improve complete trajectories, but they usually do not explicitly identify planning actions or measure their importance for task success. To address this issue, we propose Planning-Aware Policy Optimization for VLA models (PAPO-VLA). PAPO-VLA first identifies planning actions by jointly considering action variation and trajectory outcome, then estimates their importance through causal sufficiency and causal necessity, and finally incorporates this importance into GRPO advantage estimation. In this way, more important planning actions receive stronger optimization emphasis, while the whole trajectory is still optimized by trajectory-level feedback. Experiments on multiple benchmarks demonstrate the effectiveness of PAPO-VLA.

Abstract

Robotic Policy Evaluation

Explore similar work

PAPO-VLA: Planning-Aware Policy Optimization for Vision-Language-Action Models

PACT-VLA: Speculative Action-Chunk Verification with Pre-Commit Observations

CheckVLA: Execution-Time Verification with Action-Conditioned World Model for Long-Horizon Mobile Manipulation