World-Calibrated Proposal-to-Action Flow for Vision-Language-Action Models

Jie He; Wei Li; Junwen Tong; Rui Shao; Wei-Shi Zheng; Liqiang Nie

cs.ROOct 1, 2026

World-Calibrated Proposal-to-Action Flow for Vision-Language-Action Models

Authors: Jie He, Wei Li, Junwen Tong, Rui Shao, Wei-Shi Zheng, Liqiang Nie

Organizations: Harbin Institute of Technology, Shenzhen · State Key Laboratory of Mobile Network and Mobile Multimedia Technology, ZTE, China · Sun Yat-sen University

Abstract

Flow-based Vision-Language-Action (VLA) policies generate action chunks by transporting samples from a task-agnostic isotropic Gaussian source. As this source is conditioned on neither recent execution nor predicted future evolution, (i) it discards the local continuity established by recently executed motion. (ii) Even when predictive world representations are introduced, they often only condition the transport dynamics rather than determine where generation starts, how far it may deviate, or along which action directions it may expand. Building on this observation, we introduce ProAct, a world-calibrated proposal-to-action framework that makes the generative source itself predictable. (i) To preserve motion continuity, a lightweight Proposal Expert converts recent actions into a scene-aware hypothesis via one motion-anchored endpoint flow-matching step, initializing generation near the demonstrated action manifold. (ii) To jointly capture intended scene evolution and proposal-future compatibility, a prospective World Expert treats the hypothesis as a soft motion prior while predicting the task-consistent latent future. (iii) From this compatibility, the model calibrates a proposal-centered anisotropic source, where a bounded per-step extent controls the allowed deviation and a trace-normalized low-rank geometry under a condition-number budget allocates refinement over coupled translation, rotation, and gripper directions. Compared with $π_{0.5}$ , ProAct improves performance across simulation and real-world tasks while reducing denoising steps by 50%, inference latency by up to 25.8%, and increasing throughput by up to 34.8%.

Figures & tables

Figure 1: Motivation of ProAct . ProAct uses recent motion to form a scene-aware proposal and prospective world reasoning to calibrate its source distribution. Starting from this informed, proposal-centered source shortens action transport while improving performance and efficiency.

Figure 2: Overview of ProAct . A Proposal Expert forms a scene-aware motion hypothesis, which the World Expert calibrates against the task-consistent future to predict refinement extent and geometry. The resulting source is then refined into executable actions by the Action Expert.

Figure 3: World-calibrated source construction. Proposal-conditioned future features yield per-step refinement extent and low-rank geometry, defining an anisotropic source for action refinement.

Method	LIBERO					LIBERO-Plus
	Spatial	Object	Goal	Long	Avg.	Camera	Robot	Language	Light	Background	Noise	Layout	Total
OpenVLA-OFT Kim et al. (2025)	97.6	98.4	97.9	94.5	97.1	56.4	31.9	79.5	88.7	93.3	75.8	74.2	69.6
$\pi_{0.5}$ Intelligence et al. (2025)	98.8	98.2	98.0	92.4	96.9	59.7	65.5	75.3	87.0	82.4	72.1	80.3	73.6
WorldVLA Cen et al. (2025)	87.6	96.2	83.4	60.0	81.8	0.1	27.9	41.6	43.7	17.1	10.9	38.0	25.0
UniVLA Bu et al. (2025)	95.4	98.8	93.6	94.0	95.4	1.8	46.2	69.6	69.0	81.0	21.2	31.9	42.9
VLA-Adapter Wang et al. (2026b)	97.8	99.2	97.2	95.0	97.3	36.2	37.9	74.6	70.6	76.1	58.0	69.7	59.1

Table 1: Results on LIBERO and LIBERO-Plus. We report performance on the four LIBERO task suites and generalization under seven perturbation dimensions in LIBERO-Plus.

Figure 4: Closed-loop proposal refinement on real robots. Drawer Toy Retrieval (left) and Block Sorting (right). Enlarged frames overlay history (purple), proposal (blue), and refined action (green); orange callouts mark changes requiring departure from local motion continuity.

Method	Block Sorting					Pot Lid Retrieval			Drawer Manipulation
	Pick Red Block	+Place Red in Bowl	+Pick Blue Block	+Place Blue in Bowl	CSR	Remove Lid	+Retrieve Object to Plate	CSR	Open Drawer	+Grasp Toy	+Place Toy Inside	+Close Drawer	CSR
OpenVLA-OFT Kim et al. (2025) †	19/25	17/25	13/25	12/25	61.0%	21/25	17/25	76.0%	17/25	15/25	12/25	12/25	56.0%
$\pi_{0}$ Black et al. (2024) †	20/25	18/25	14/25	14/25	66.0%	21/25	16/25	74.0%	18/25	17/25	13/25	13/25	61.0%
$\pi_{0.5}$ Intelligence et al. (2025) †	23/25	22/25	19/25	17/25	81.0%	22/25	20/25	84.0%	21/25	21/25	18/25	16/25	76.0%
ProAct	23/25	23/25	21/25	20/25	87.0%	24/25	21/25	90.0%	23/25	22/25	20/25	19/25	84.0%

Table 3: Real-world manipulation results. Results are reported over the same six tasks on both on GALAXEA R1 Lite Galaxea (2025) and AgileX Cobot Magic Fu et al. (2024) platforms. The superscript

\dagger

denotes our reproduced baselines.

Table 7

Figure 7: World-calibrated proposal refinement. Left: proposal refinement in 3D translation space. Right: normalized trajectories and calibrated support across seven action dimensions.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Optimization
Setting	LIBERO	RoboTwin 2.0	Real-World
Training steps	60k	30k	30k
Batch size	128	64	64
Optimizer	AdamW, $\beta_{1}=0.9$ , $\beta_{2}=0.95$
Peak learning rate	$5\times 10^{-5}$	$5\times 10^{-5}$	$5\times 10^{-5}$
Warm-up steps	10k	10k	10k

Appendix

Table 8: Training Configuration of ProAct .

Method	Adjust [-1pt]Bottle	Beat [-1pt]Block [-1pt]Hammer	Blocks [-1pt]Rank. [-1pt]RGB	Blocks [-1pt]Rank. [-1pt]Size	Click [-1pt]Alarmclock	Click [-1pt]Bell	Dump [-1pt]Bin [-1pt]Bigbin	Grab [-1pt]Roller	Handover [-1pt]Block	Handover [-1pt]Mic
DP3 Ze et al. (2024)	3%	8%	0%	0%	14%	0%	53%	2%	0%	3%
$\pi_{0}$ Black et al. (2024)	56%	21%	5%	1%	11%	3%	24%	80%	8%	13%
$\pi_{0.5}$ Intelligence et al. (2025)	83%	50%	35%	14%	93%	76%	42%	89%	19%	18%
X-VLA Zheng et al. (2026)	33%	9%	4%	25%	43%	63%	41%	20%	1.3%	1%
Spatial Forcing Li et al. (2026a)	71%	13%	7%	3%	28%	28%	70%	94%	2%	2%
X-WAM Guo et al. (2026)	13%	21%	36%	0%	61%	52%	40%	16%	21%	8%

Appendix

Table 9: Per-task success rates on 50 RoboTwin 2.0 tasks under Randomized (Hard). All methods use the same protocol. The overall average is reported in the final block.

Figure 8: Representative successful real-world rollouts of ProAct on the six evaluation tasks. Within each task panel, the top row shows GALAXEA R1 Lite and the bottom row shows AgileX Cobot Magic; snapshots in each row are ordered from left to right. The eight snapshots are selected at semantic milestones rather than at uniform temporal intervals, covering repeated sorting, container and drawer manipulation, object transfer, and sustained wiping.

Figure 9: Per-checkpoint real-world ablations across the six GALAXEA R1 Lite tasks. Each panel plots per-checkpoint success for one ablation axis: motion proposal (left), world calibration (middle), and source calibration (right). Within every panel the full ProAct (blue) attains the highest bar, and its margin widens at the later transition and terminal checkpoints where local motion continuity becomes unreliable across the full execution horizon .

Method	Block Sorting					Pot Lid Retrieval			Drawer Manipulation
Method	Pick Red Block	+Place Red in Bowl	+Pick Blue Block	+Place Blue in Bowl	CSR	Remove Lid	+Retrieve Object	CSR	Open Drawer	+Grasp Toy	+Place Toy Inside	+Close Drawer	CSR
OpenVLA-OFT Kim et al. (2025) †	17/25	15/25	12/25	11/25	55.0%	20/25	15/25	70.0%	16/25	15/25	11/25	10/25	52.0%
$\pi_{0}$ Black et al. (2024) †	19/25	17/25	14/25	13/25	63.0%	20/25	14/25	68.0%	18/25	16/25	13/25	12/25	59.0%
$\pi_{0.5}$ Intelligence et al. (2025) †	21/25	19/25	17/25	16/25	73.0%	21/25	18/25	78.0%	20/25	19/25	16/25	15/25	70.0%
ProAct	22/25	21/25	20/25	19/25	82.0%	23/25	20/25	86.0%	22/25	21/25	19/25	18/25	80.0%

Appendix

Table 10: Per-checkpoint real-world results on AgileX Cobot Magic. For each of the six tasks we report the number of successful trials (out of 25) at every progressive checkpoint, so that a checkpoint is credited only when all preceding stages succeed. CSR is the mean success rate across a task’s checkpoints, and its final checkpoint equals the task TSR. Best entries per column are in bold .

Figure 10: Additional world-calibrated proposal refinements. Further examples in the same format as Fig. 7 , illustrating refinement behavior across more cases and execution phases.

World Models for VLA Models

See all 28 papers

CF-VLA: Efficient Coarse-to-Fine Action Generation for Vision-Language-Action Policies
Apr 27, 2026Fan Du, Feng Yan, Jianxiong Wu +8Robotic Control Efficient VLA Model Inference
1Southern University of Science and Technology · 2Xi’an Jiaotong University · 3United Nova Technology +1
TokenPB: Enabling VLA Action Self-Verification via VLM Token Probability Bucketing
Mar 18, 2026Chen Zhao, Zhuoran Wang, Haoyang Li +6Vision-Language-Action Models VLA Model Training
School of Information, Renmin University of China · Key Laboratory of Data Engineering and Knowledge Engineering, Beijing, China · Beijing University of Posts and Telecommunications +1
World Guidance: World Modeling in Condition Space for Action Generation
Feb 25, 2026Yue Su, Sijin Chen, Haixin Shi +7Vision-Language-Action Models World Model Learning
ByteDance Seed · The University of Hong Kong

Apr 27, 2026J/K move · Enter open · S save

CF-VLA: Efficient Coarse-to-Fine Action Generation for Vision-Language-Action Policies

Fan Du, Feng Yan, Jianxiong Wu, Xinrun Xu+7

1Southern University of Science and Technology · 2Xi’an Jiaotong University · 3United Nova Technology+1

Robotic Control Efficient VLA Model Inference Visuomotor Policy Learning Efficient VLA Models Robot Motion Generation Vision-Language-Action Models+2

Open paper AskPDF arXiv

Flow-based vision-language-action (VLA) policies offer strong expressivity for action generation, but suffer from a fundamental inefficiency: multi-step inference is required to recover action structure from uninformative Gaussian noise, leading to a poor efficiency-quality trade-off under real-time constraints. We address this issue by rethinking the role of the starting point in generative action modeling. Instead of shortening the sampling trajectory, we propose CF-VLA, a coarse-to-fine two-stage formulation that restructures action generation into a coarse initialization step that constructs an action-aware starting point, followed by a single-step local refinement that corrects residual errors. Concretely, the coarse stage learns a conditional posterior over endpoint velocity to transform Gaussian noise into a structured initialization, while the fine stage performs a fixed-time refinement from this initialization. To stabilize training, we introduce a stepwise strategy that first learns a controlled coarse predictor and then performs joint optimization. Experiments on CALVIN and LIBERO show that our method establishes a strong efficiency-performance frontier under low-NFE (Number of Function Evaluations) regimes: it consistently outperforms existing NFE=2 methods, matches or surpasses the NFE=10

π_{0.5}

baseline on several metrics, reduces action sampling latency by 75.4%, and achieves the best average real-robot success rate of 83.0%, outperforming MIP by 19.5 points and

π_{0.5}

by 4.0 points. These results suggest that structured, coarse-to-fine generation enables both strong performance and efficient inference. Our code is available at https://github.com/EmbodiedAI-RoboTron/CF-VLA.

Abstract

World Models for VLA Models

Explore similar work

CF-VLA: Efficient Coarse-to-Fine Action Generation for Vision-Language-Action Policies

TokenPB: Enabling VLA Action Self-Verification via VLM Token Probability Bucketing

World Guidance: World Modeling in Condition Space for Action Generation