StreamingVLA: Streaming Vision-Language-Action Model with Action Flow Matching and Adaptive Early Observation

Yiran Shi; Dongqi Guo; Tianchen Zhao; Feng Gao; Liangzhi Shi; Chao Yu; ZhiJian Mo; Qihua Xiao

cs.ROMar 30, 2026

StreamingVLA: Streaming Vision-Language-Action Model with Action Flow Matching and Adaptive Early Observation

Authors: Yiran Shi, Dongqi Guo, Tianchen Zhao, Feng Gao, Liangzhi Shi, Chao Yu, ZhiJian Mo, Qihua Xiao, +3 more

Organizations: Tsinghua University · Lenovo Group Ltd.

Abstract

Vision-language-action (VLA) models have demonstrated exceptional performance in natural language-driven perception and control. However, the high computational cost of VLA models poses significant efficiency challenges, particularly for resource-constrained edge platforms in real-world deployments. However, since different stages of VLA (observation, action generation and execution) must proceed sequentially, and wait for the completion of the preceding stage, the system suffers from frequent halting and high latency. To address this, We conduct a systematic analysis to identify the challenges for fast and fluent generation, and propose enabling VLAs with the ability to asynchronously parallelize across VLA stages in a "streaming" manner. First, we eliminate the reliance on action chunking and adopt action flow matching, which learns the trajectory of action flows rather than denoising chunk-wise actions. It overlaps the latency of action generation and execution. Second, we design an action saliency-aware adaptive observation mechanism, thereby overlapping the latency of execution and observation. Without sacrificing performance, StreamingVLA achieves substantial speedup and improves the fluency of execution. It achieves a 2.4 $\times$ latency speedup and reduces execution halting by 6.5 $\times$ .

Figures & tables

Figure 1: StreamingVLA explores streaming execution of VLA models by enabling different stage to be runned in a asynchronous manner. By overlapping the latency for stages, without sacrificing performance, it achieves 2.4 $\times$ end-to-end speedup, and 6.5 $\times$ execution halting time reduction, enabling fast and fluent execution.

Figure 2: The overall methodology framework for StreamingVLA: We conduct a systematic timeline analysis and conclude the optimization target for fast and fluent execution. We present two key techniques: action flow matching and adaptive early observation, to overlap the latency of action execution with of action generation and VLM observation, respectively.

Figure 3: The Illustration of state-based modeling of action flow matching: It reformulates the action modeling from predicting actions as absolute values to predicting actions as updates to a feature-space state accumulated by prior actions. Extended formulation and architectural adjustment is adopted for larger scaled VLA model and mainstream benchmarks.

Figure 4: The Illustration of action saliency aware adaptive early observation: We highlight the importance of diverse action saliency, and adopt a lightweight predictor to implement an adaptive early observation scheme.

Method	Success Rate (%) $\uparrow$					Time per Action (ms) $\downarrow$	Halting Gap (ms) $\downarrow$
Method	Spatial	Object	Goal	Long	Average	$T_{\text{action}}$	$T_{\text{halt}}$
$\pi_{0.5}$ ( $h=5$ )	98.8	98.2	98.0	92.4	96.9	74.5	232.3
$\pi_{0.5}$ ( $h=10$ )	97.4	98.2	96.2	88.6	95.1	49.9 ( 1.49 $\times$ )	230.8 ( 1.01 $\times$ )
RTC (d=1) [ 4 ]	96.2	19.8	93.0	25.2	58.55	50.2 ( 1.48 $\times$ )	203.6 ( 1.14 $\times$ )
SmolVLA [ 33 ]	97.0	99.0	97.0	90.0	95.8	51.1 ( 1.45 $\times$ )	180.7 ( 1.28 $\times$ )
VLASH [ 35 ]	97.5	99.2	97.3	94.6	97.1	40.6 ( 1.83 $\times$ )	-

Table 1: Main results on the LIBERO benchmark. We present the performance and efficiency of StreamingVLA and baseline methods. The

T_{\text{action}}

and

T_{\text{halt}}

are adopted to measure the speed and fluency of execution. The “AFM” and “AEO” represent the “action flow matching” and “adaptive early observation” technique. The “NEO” stands for the “naive early observation”. The “ANAO” stands for the “action norm based adaptive observation” as mentioned in Sec. 4.2

AFM	NM	SA	Horizon	EO	Success Rate(%)	$T_{\text{action}}$ (ms)	$T_{\text{halt}}$ (ms)
-	-	-	10	-	95.1	49.9	230.78
✓	-	-	10	-	-	-	-
✓	✓	-	10	-	61.8	-	-
✓	✓	✓	10	-	97.1	33.7	76.1
✓	✓	✓	10	Naive (2)	86.2	29.3	23.0
✓	✓	✓	10	Random (1.4)	90.875	32.125	38.3

Table 2: Ablation studies for StreamingVLA techniques validate the effectiveness of our Action Flow Matching(AFM) formulation w/o state-based alignment modeling(SA), normalization modification(NM) mentioned in Sec. 4.1 , as well as adaptive early observation(EO) schemes. The “2” and “1.45” represents the average number of early observed actions. The "Random" represents employing early observation randomly.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Real-world Setup. The real-world evaluation platform includes a tabletop workspace, a Franka Panda robotic manipulator fixed to the table, and a RGB camera for visual observation. All objects are positioned on the table surface.

Figure 6: The runtime breakdown based on actual profiling of $\pi_{0.5}$ and StreamingVLA. The three colored blocks correspond to the “observation”, "gen-action" and "execution" stage in the VLA workflow. The length of each block reflects the measured time.

Figure 7: Architecture of the predictor for adaptive early observation. During training, the model receives the early-frame embedding and the pending action sequence as conditions and predicts a residual embedding $\Delta embedding$ that estimates the latent change toward the future frame. The model is trained using an MSE loss between the ground-truth late-frame embedding and the reconstructed representation ( $embedding_{early}+\Delta embedding$ ). At inference time, the $L_{2}$ norm of $\Delta embedding$ serves as a visual-variance indicator, which is compared with a threshold to determine whether early observation should be triggered.

Figure 8: Screenshots of our video simulation results. The images in each row are spaced 4 seconds apart. Top two rows: open the top drawer and put the bowl inside. StreamingVLA completes the task in under 16 seconds, whereas pi0.5 has not yet picked up the bowl at the 20-second mark. Bottom two rows: put both the alphabet soup and the tomato sauce in the basket. At 20 seconds, StreamingVLA has already finished the pick-and-place operations for both objects, while pi0.5 is still on its way to picking up the second item.

Explore similar work

FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference
Aug 27, 2026Zekai Li, Jiaming Tang, Zhijian LiuEfficient VLA Model Inference VLMs for Robotics
UC San Diego · MIT
AdaVLA: Adaptive Step Flow Matching for Training-free Acceleration of Vision-Language-Action Models
Aug 29, 2026Sunghwan Han, Youngtae Han, Youngmin YiFlow Matching Vision-Language-Action Models
Dept. of Artificial Intelligence, Sogang University, Seoul, 04107, Republic of Korea
TokenPB: Enabling VLA Action Self-Verification via VLM Token Probability Bucketing
Mar 18, 2026Chen Zhao, Zhuoran Wang, Haoyang Li +6Vision-Language-Action Models
School of Information, Renmin University of China · Key Laboratory of Data Engineering and Knowledge Engineering, Beijing, China · Beijing University of Posts and Telecommunications +1

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

FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference

Zekai Li, Jiaming Tang, Zhijian Liu

UC San Diego · MIT

Efficient VLA Model Inference VLMs for Robotics Action Chunking

Open paper AskPDF arXiv

Vision-Language-Action (VLA) models are increasingly promising for robotic manipulation, yet their real-world deployment remains bottlenecked by high inference latency and unstable asynchronous execution. This challenge is particularly pronounced in flow-matching-based VLA models, where action decoding requires multiple iterative steps conditioned on the VLM context. While efficient inference methods improve control frequency and asynchronous methods reduce execution idle time, existing approaches often fail to jointly achieve low-latency inference and accurate, temporally consistent asynchronous execution. We introduce FlashVLA, a streaming action decoding framework that addresses both challenges in a unified formulation. FlashVLA maintains a streaming action buffer with multiple chunks at different noise levels and decodes them using chunk-wise causal attention. This design allows FlashVLA to produce one executable action chunk per inference step. Moreover, its chunk-wise autoregressive formulation implicitly preserves action continuity, enabling smooth asynchronous execution without extra future-state conditioning. Across extensive simulated and real-world experiments, FlashVLA substantially improves inference speed while maintaining strong task performance. It can achieve

\geq

30,Hz control frequency on a single GPU with smooth asynchronous inference in real-world deployment.

Abstract

Explore similar work

FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference

AdaVLA: Adaptive Step Flow Matching for Training-free Acceleration of Vision-Language-Action Models

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