CommitFlow: Semantic Commitment Verification and Local Correction for Long-Horizon Robot Manipulation VLA Execution

Zixiang Zhao; Yansong Feng; Yang Yang; Chaoyu Wang; Haoran Xiao; Hui Zhang; Chuang Cheng; Jianjun Ma

cs.ROSep 18, 2026

CommitFlow: Semantic Commitment Verification and Local Correction for Long-Horizon Robot Manipulation VLA Execution

Authors: Zixiang Zhao, Yansong Feng, Yang Yang, Chaoyu Wang, Haoran Xiao, Hui Zhang, Chuang Cheng, Jianjun Ma

Organizations: National University of Defense Technology, Changsha, China

Abstract

Although vision-language-action (VLA) policies have advanced rapidly, long-horizon execution may still progress to the next task stage before the required physical effect has been established. We call this a mismatch between semantic commitments, physical conditions that a stage must establish or maintain, and the actual physical state. Because an action command alone cannot confirm such a condition, local deviations can propagate and cause task failure. To address this problem, we present CommitFlow, a closed-loop execution framework that combines commitment monitoring with local correction while keeping the base policy frozen. CommitFlow integrates three components. A Semantic Commitment Monitor (SCM) compares stage requirements against current state evidence and holds back dependent actions when a required condition is unmet or violated. BoundaryFlow then generates a local correction conditioned on the current state and base action, and Relation and Gain Calibration (RGC) selects the smallest correction strength that satisfies the relevant constraints. Across the ten common RoboTwin 2.0 benchmark tasks, CommitFlow achieves a mean success rate of 75.9 percent, improving on the base policy pi0.5 by 22.7 percent. Cross-policy experiments show consistent gains, pointing toward reliable long-horizon robot execution.

Figures & tables

Fig. 2: CommitFlow overview. RGB-D encoding supplies visual features, metric relations, and task context. SCM assigns action permissions; BoundaryFlow predicts a base-action-conditioned residual; RGC selects its strength. The permission-gated action is executed, and fresh observations support re-verification.

Fig. 3: BoundaryFlow action generation and local correction. The frozen VLA generates a base action chunk through flow matching, while the trainable BoundaryFlow branch generates a corrective action chunk. Combining the two produces the corrected action chunk.

Fig. 4: RGB-D/CAD localization and SCM recognition performance. (a) Median (P50) and 90th-percentile (P90) position errors in XY and XYZ over all 1,971 closed-loop frames and the 259-frame ALIGN TO GRASP subset. (b) Macro-F1 for stage and event-class recognition across six streams (1,971 frames). The dashed line marks the 95% target.

Method	Handover Block	Handover Mic	Hang Mug	Bread Basket	Bread Skillet	Place A2B Left	Place A2B Right	Object Basket	Object Stand	Object Scale	Object Cabinet	Avg. 10
RDT	45	90	1	10	5	3	1	33	15	1	33	23.6
$\pi_{0.5}$	24	58	11	62	62	51	39	69	68	45	54	53.2
$\pi_{0.5}$ +TACO [ 14 ]	36	63	–	65	61	56	42	78	78	52	56	58.7
$\pi_{0.5}$ +CAPS [ 15 ]	38	72	–	63	69	54	47	76	81	57	58	61.5
$\pi_{0.5}$ +CommitFlow	66	60	42	75	78	70	69	88	97	85	71	75.9

TABLE I: RoboTwin 2.0 Results (%)

Fig. 5: Qualitative execution sequences in simulation. Four examples compare the base policy (upper rows) with CommitFlow (lower rows). Frames progress from left to right; marked regions highlight the relevant physical relations.

Fig. 6: Recovery-strength sweep. Cabinet and Basket: overall task success (left axis), recovery-error reduction and zero-crossing overshoot in mm (right axis). Circled annotations highlight selected measurements; the success axis is truncated.

Fig. 7: Qualitative execution sequences on the real robot. Bowl stacking (left) and eraser placement (right), comparing the base policy (upper row) with CommitFlow (lower row). Frames progress from left to right.

Task	$\pi_{0}$ [ 20 ]	$\pi_{0.5}$ [ 19 ]	LingBot-VLA [ 21 ]
Handover Block	26 $\rightarrow$ 45	24 $\rightarrow$ 66	85 $\rightarrow$ 90
Handover Mic	86 $\rightarrow$ 89	58 $\rightarrow$ 60	95 $\rightarrow$ 96
Hang Mug	3 $\rightarrow$ 25	11 $\rightarrow$ 42	49 $\rightarrow$ 62
Bread Basket	35 $\rightarrow$ 50	62 $\rightarrow$ 75	89 $\rightarrow$ 91
Bread Skillet	12 $\rightarrow$ 23	62 $\rightarrow$ 78	80 $\rightarrow$ 83
Place A2B Left	30 $\rightarrow$ 45	51 $\rightarrow$ 70	83 $\rightarrow$ 83

TABLE II: Cross-Policy Results (%)

Task	Fresh Base	Fresh + IK	Fresh + Ours
Object Cabinet	61	57	71
Object Basket	60	68	88
Place A2B Left	55	57	70
Place A2B Right	56	60	69
Object Scale	78	78	85
Object Stand	91	94	97

TABLE III: Local Correction Ablation (%)

Correction reuse — simulation
Task	Base	+ CommitFlow
Handover Mic	58	65
Can $\rightarrow$ Basket	40	58
Real-world execution
Eraser $\rightarrow$ Container	50	63
Stack Three Bowls	40	50

TABLE IV: Correction Reuse and Real-World Results (%)

Long-Horizon Robotic Manipulation

Long-horizon robotic manipulation studies how robots plan, remember, and recover across extended sequences of physical tasks.

See all 435 papers

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
VLA-Corrector: Stage-Aware Observable State Understanding for Prompt-Based Closed-Loop Recovery of Vision-Language-Action Policies
Sep 6, 2026Chang Song, Bin Qian, Yan Feng +1Long-Horizon Robotic Manipulation Robustness of VLA Models
TemporalFlow-VLA: Learning Physically Grounded Execution History for Long-Horizon Robot Manipulation
Aug 27, 2026Jiarui Yang, Yehao Lu, Yuning Su +9Robot Learning Long-Horizon Robotic Manipulation
The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China. · Zhejiang University, Hangzhou, China. · Simon Fraser University, Burnaby, BC, Canada. +1

Jul 29, 2026J/K move · Enter open · S save

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

Yushan Liu, Peibo Sun, Xintao Chao, Zhenyang Yang+7

Tsinghua University · Shanghai Jiao Tong University · Peking University+2

Long-Horizon Robotic Manipulation World Models for Robotics Vision-Language-Action Models Runtime Verification

Open paper AskPDF arXiv

Vision-language-action (VLA) policies commonly execute long-horizon mobile manipulation through open-loop action chunks, issuing multiple actions without receiving new high-level visual input. A committed chunk therefore implies how observations should evolve, but accidental deviations can violate this expectation while the remaining actions continue to propagate the error: commit-time policy confidence cannot react to a deviation that occurs after dispatch, and observation-only anomaly scores lack an action-conditioned reference for separating expected effects from unexplained changes. We propose CheckVLA, which verifies execution with a separately trained, frozen action-conditioned world model. A conformally calibrated risk threshold bounds the episode-level probability of an unnecessary first intervention and determines when to intervene, its exceedance controls how strongly the rewritten suffix retains the superseded chunk, latency-aware hard prefixing restricts replacement to actions that remain deployable, and an event-driven keyframe bank preserves evidence of prior progress across repairs. On RoboCasa365, under a common training recipe and a matched invocation budget, CheckVLA attains a 36.1% average success rate against 27.6% for periodic replanning (+8.5 points). At a matched 5% episode-level false-alarm target, action conditioning raises timely recall to 77.9%, against 48.6% for an observation-only control and 37.9% for an action-shuffled control. These simulation results support action-conditioned verification as a way to restore feedback during chunked execution while keeping the repair consistent with inference latency.

Abstract

Long-Horizon Robotic Manipulation

Explore similar work

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

VLA-Corrector: Stage-Aware Observable State Understanding for Prompt-Based Closed-Loop Recovery of Vision-Language-Action Policies

TemporalFlow-VLA: Learning Physically Grounded Execution History for Long-Horizon Robot Manipulation