PRISM: Precision Forcing for Generalist Vision-Language-Action Policies

Chenghao Yue; Peiyan Li; Hao Wang; Lei Lv; Jiahao Chen; Huaping Liu; Yu Luo; Yuanlong Zhang

cs.ROOct 10, 2026

PRISM: Precision Forcing for Generalist Vision-Language-Action Policies

Authors: Chenghao Yue, Peiyan Li, Hao Wang, Lei Lv, Jiahao Chen, Huaping Liu, Yu Luo, Yuanlong Zhang

Organizations: Tsinghua University · Dalian University of Technology · Tongji University · Peking University ·

Abstract

Coupling a pretrained vision-language model (VLM) with a diffusion or flow-matching action expert has become the mainstream vision-language-action (VLA) paradigm for embodied policies, yet they remain unreliable once a task demands precise manipulation, e.g., submillimeter cable insertion. We attribute this weakness to a training objective that averages flow-matching errors across contexts and flow times, allocating little supervision to the few decisive moments determining task success. Along the generative path, inference integrates a locally supervised velocity field, so equally penalized errors at different flow times can affect the executed action unequally. Consequently, precision-task performance deteriorates as physical tolerance tightens by orders of magnitude. We call this supervision imbalance across contexts and flow times Flow-Time Starvation. The remedy is to jointly align supervision allocation and representation learning with context-dependent precision demand, enabling the policy to recognize when greater precision is needed. We instantiate this precision-forcing strategy as Precision-aware Reweighting and Internalization for Skilled Manipulation (PRISM), which scores precision demand offline, strengthens flow supervision where errors are consequential and action signals are visible, and internalizes that demand in the VLM condition representation. Simulation and real-robot experiments show that PRISM preserves or improves general-task performance while raising mean end-to-end success from 38.3% to 86.7% across drive connector insertion, bimanual connector mating, and screw tightening, substantially extending the precision boundary of existing embodied policies without new sensors or observation modalities.

Figures & tables

Figure 1: Overview of Prism . PPS estimates precision demand from robot trajectories. CPR learns precision-aware VLM features, while PSW focuses action supervision on precision-demanding contexts with visible action evidence, improving fine manipulation without changing policy inference.

Figure 2: Why precise manipulation needs different supervision. Left, top to bottom: key frames and tolerances for cable pickup, bimanual handover, and connector insertion, the three stages of the real-robot submillimeter drive-insertion task. (a) Similar action errors incur different tolerance-normalized burdens. (b) Inverse-tolerance-weighted endpoint responses to equal-energy velocity perturbations peak near the clean endpoint for fine manipulation.

Figure 3: Overview of Prism . (a) PPS combines joint motion, end-effector motion, and gripper engagement into offline precision demand. (b) CPR predicts this demand from mask-pooled VLM prefix features, while PSW reweights the flow-matching loss using offline demand and flow-time visibility derived from action RMS and flow time. Both objectives leave inference unchanged.

Method	Close box	Close laptop	Toilet seat	Sweep	Close fridge	Phone	Umbrella	Frame	Wine	Water plants	Mean
OpenVLA	0.60	0.35	0.75	0.55	0.85	0.20	0.30	0.15	0.20	0.05	0.40
SpatialVLA	0.80	0.70	0.85	0.20	0.80	0.15	0.25	0.40	0.15	0.30	0.46
$\pi_{0}$	0.90	0.60	1.00	0.30	0.90	0.25	0.35	0.75	0.05	0.45	0.55
CogACT	0.90	0.80	0.95	0.50	0.85	0.50	0.55	0.45	0.30	0.25	0.61
CoT-VLA	0.95	0.75	1.00	0.80	0.65	0.50	0.40	0.50	0.55	0.50	0.66
HybridVLA	0.85	0.95	1.00	0.90	1.00	0.50	0.50	0.70	0.50	0.50	0.74

Table 1: RLBench-10 success rates (absolute gains in parentheses). Green highlights Prism .

Figure 4: Representative stages of three MSSM tasks, ordered temporally from left to right.

Method	Drive connector insertion (0.15 mm)					Bimanual connector mating (0.1 mm)							Screw tightening (0.2 mm)					Avg.
	S1	S2	S3	SR $\uparrow$	$T\downarrow$	S1	S2	S3	S4	S5	SR $\uparrow$	$T\downarrow$	S1	S2	S3	SR $\uparrow$	$T\downarrow$	SR $\uparrow$
$\pi_{0.5}$	1.00	0.90	0.35	0.35	47	0.95	0.95	0.50	0.55	0.65	0.35	65	0.95	0.85	0.45	0.45	37	0.38
$\pi_{0.5}$ + DAgger	1.00	1.00	0.65	0.65	43	0.95	1.00	0.70	0.75	0.85	0.60	59	1.00	0.95	0.70	0.70	35	0.65
$\pi_{0.5}$ + Prism	1.00	1.00	0.95	0.95	38	1.00	1.00	0.85	0.90	0.95	0.80	52	1.00	1.00	0.85	0.85	32	0.87

Table 3: MSSM real-robot results. S1–S5 report stage success. SR is end-to-end task success.

T

is mean completion time in seconds over successful trials. Avg. averages SR over the three tasks.

Table 7

Panel of Figure 5 — Figure 5: Mechanism analysis of precision forcing. (a) CPR predictions versus offline PPS on held-out insertion trajectories. (b) Residual reduction over $\pi_{0.5}$ across five equal-count PPS bins. (c) Residual profiles in four equal-count SNR groups, formed per flow time within high-PPS frames. Purple/teal indicates a decrease/increase under PSW. (d) Relative CPR-conditioned visual sensitivity across fine-manipulation contexts.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	StarVLA	$\pi_{0.5}$
VLM backbone	Qwen3-VL-4B-Instruct	PaliGemma (Gemma 2B)
Action expert	QwenPI_v3, 36 layers, width 1,024	Gemma 300M
Input image resolution	$224\times 224$	$224\times 224$
Robot state input	Yes	Yes
Action horizon $H$	15	15
Action dimensions (dataset-specific)	8 (single-arm RLBench)	16 (MSSM); padded to 32

Appendix

Table 6: Training configurations, real-robot execution settings, and offline PPS efficiency. PPS measurements use insertion trajectories and are independent of the policy backbone.

Task	Objects and specifications
Drive connector insertion	Geek+ GS-SD-48V-20A-21D V01.01 dual-channel safety motor drive and a Hirose DF22 cable assembly with a matching drive receptacle.
Bimanual connector mating	Three connector pairs, used in order: white Hirose DF22 2-pin (0.2 mm tolerance), white Hirose DF22 4-pin (0.1 mm), and black Hirose DF60 (0.3 mm). Male halves start on the left and female halves on the right.
Screw tightening	Bosch GO Gen 2 cordless screwdriver with a compatible bit; a loose screw pre-positioned in the lower-left threaded hole of the motor-drive housing. Measured alignment tolerance: 0.2 mm.

Appendix

Table 7: Objects used in the three MSSM tasks.

Task	Stage	Description
Drive connector insertion	S1: Cable pickup	The left arm picks up the cable behind its connector housing.
	S2: Bimanual handover	The left arm transfers the cable to the right arm.
	S3: Connector insertion	The right arm aligns the cable connector with the receptacle and inserts it.
Bimanual connector mating	S1: Left-arm pickup	The left arm grasps the male half of the first connector pair.
	S2: Right-arm pickup	The right arm grasps the female half of the first connector pair.
	S3: First pair	Align and mate the Hirose DF22 2-pin pair, then place it back on the workbench.

Appendix

Table 8: MSSM tasks, stages, and operation descriptions.

Figure 8: CPR sensitivity across three tasks and three camera views. Each column shows one task. Rows show the head, left-wrist, and right-wrist cameras at the same selected observation within that task. Labels identify the action and camera. Colors indicate relative within-view sensitivity, not a shared absolute response scale. Selected views and heatmaps are jointly rotated for readability.

Figure 9: PSW amplification on drive connector insertion. Success peaks at

\gamma=3

Figure 10: Equal-energy perturbations across flow time. Curves show unweighted endpoint action MSE per unit perturbation energy on ten held-out episodes. Shaded bands are 95% episode-bootstrap intervals. The clean-end interval produces the largest response in each operation stage at both energy settings.

Vision-Language-Action Models

Vision-language-action models turn visual observations and natural-language instructions into actions for robots and other embodied agents.

See all 1,190 papers

Sep 24, 2026

ActGaze: Learning Action-Grounded Gaze through Counterfactual Visual Interventions for High-Precision Manipulation

Current Vision-Language-Action (VLA) models often struggle with high-precision robotic manipulation. We attribute this limitation primarily to their visual attention being dispersed across task-irrelevant regions. To address this issue, we propose ActGaze, a training approach that guides VLA policies to gaze on task-relevant regions, much like humans gaze on critical visual cues while executing precise movements. Unlike prior methods that rely on external labels for gaze supervision, ActGaze derives spatial supervision directly from the VLA's own action objective by using counterfactual visual interventions to identify regions that are critical for action prediction. Extensive real-robot experiments on four high-precision robotic manipulation tasks demonstrate that ActGaze induces more focused visual attention on task-relevant regions and consistently outperforms the base VLA policy and other visual-grounding approaches.

Jinxuan Zhu, Jiaheng Wang, Chao Tang +7

Samsung Robotics eXperience · National University of Singapore

Visuomotor Policy Learning Robotic Manipulation Vision-Language-Action Models

Ask PDF

Apr 20, 2026

AnchorRefine: Synergy-Manipulation Based on Trajectory Anchor and Residual Refinement for Vision-Language-Action Models

Precision-critical manipulation requires both global trajectory organization and local execution correction, yet most vision-language-action (VLA) policies generate actions within a single unified space. This monolithic formulation forces macro-level transport and micro-level refinement to be optimized under the same objective, causing large motions to dominate learning while suppressing small but failure-critical corrective signals. In contrast, human manipulation is structured by global movement planning together with continuous local adjustment during execution. Motivated by this principle, we propose AnchorRefine, a hierarchical framework that factorizes VLA action modeling into trajectory anchor and residual refinement. The anchor planner predicts a coarse motion scaffold, while the refinement module corrects execution-level deviations to improve geometric and contact precision. We further introduce a decision-aware gripper refinement mechanism to better capture the discrete and boundary-sensitive nature of gripper control. Experiments on LIBERO, CALVIN, and real-robot tasks demonstrate that AnchorRefine consistently improves both regression-based and diffusion-based VLA backbones, yielding gains of up to 7.8% in simulation success rate and 18% in real-world success rate.

Tingzheng Jia, Kan Guo, Lanping Qian +6

Beijing University of Technology Beijing, China · Beihang University Beijing, China · Nanjing University of Science and Technology Nanjing, China

Robot Policy Learning Language-Conditioned Robot Manipulation Robotic Manipulation Vision-Language-Action Models Residual Policy Learning

Ask PDF

Jun 3, 2026

Potential-Guided Flow Matching for Vision-Language-Action Policy Improvement

Large vision-language-action (VLA) policies are increasingly trained as conditional generative models over action chunks. Yet deployment produces mixed-quality experience-successful demonstrations, partial completions, recoverable mistakes, and failures-that is difficult to use with standard imitation. Full behavior cloning (BC) imitates failures, filtered BC discards useful sub-trajectories, and offline reinforcement learning adds a large critic. We introduce ForesightFlow, a self-guided flow-matching policy that augments each generated action chunk with a learned success-potential trajectory. The same flow proposes and scores candidate actions, enabling best-of-

K

inference without an external critic. The key issue is that policy improvement and value calibration require different supervision: advantage weighting should emphasize high-quality actions, but applying the same weights to potential coordinates suppresses failure gradients and creates overconfident scores. We address this with decoupled advantage-weighted flow matching, applying exponentiated advantage weights only to action velocities while training potential velocities uniformly. We further derive a one-step boundary estimator for conditional flow matching, allowing advantage computation with a single stop-gradient forward pass. Across five BEHAVIOR-1K simulation tasks and five real-world bimanual tasks, ForesightFlow improves over imitation baselines, matches the strongest separate-critic baseline in simulation success, improves real-world success, and reduces training compute by

38\%

. Ablations show that decoupling prevents value hallucination, the one-step estimator preserves candidate-ranking fidelity, and self-guided sampling improves long-horizon execution.

Yunpeng Mei, Jiakai He, Hongjie Cao +12

1Beijing Institute of Technology · 2Tsinghua University · 3The Hong Kong University of Science and Technology (Guangzhou)

Flow Matching Visuomotor Policy Learning Vision-Language-Action Models Flow-Based Generative Modeling Flow-Based VLA Models Flow Matching for Action Generation

Ask PDF

Abstract

Vision-Language-Action Models

Explore similar work

ActGaze: Learning Action-Grounded Gaze through Counterfactual Visual Interventions for High-Precision Manipulation

AnchorRefine: Synergy-Manipulation Based on Trajectory Anchor and Residual Refinement for Vision-Language-Action Models

Potential-Guided Flow Matching for Vision-Language-Action Policy Improvement