FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference

Zekai Li; Jiaming Tang; Zhijian Liu

cs.ROAug 27, 2026

FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference

Authors: Zekai Li, Jiaming Tang, Zhijian Liu

Organizations: UC San Diego · MIT

Abstract

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.

Figures & tables

Panel of Figure 1 — Figure 1: (a) Two prevailing VLA inference patterns and FlashVLA’s joint solution. (b) Per-policy-invocation profile of $\pi_{0.5}$ and FlashVLA. Action decoding is the dominant bottleneck.

Panel of Figure 2 — Figure 2: Overview of FlashVLA. Left : Multi-buffer joint fine-tuning packs all cold-start configurations for one observation into a single sample. Attention masks isolate buffers, and chunk-wise causal attention lets noisier chunks attend only to cleaner ones. Right : Streaming inference runs the buffer as a queue: cold-start fills it with $N{-}1$ warm-up passes, while steady streaming advances all chunks, executes the cleanest one, and appends fresh noise. Gray squares denote padding.

Method	Success Rate (%)					Efficiency
	Spatial	Object	Goal	Long	Average	Steps	Time (s)	Time / Step (ms)
$\pi_{0.5}$	98.8	98.2	98.0	92.4	96.9	156.0	8.4	53.8
+ VLASH (d=1)	98.8	99.2	96.7	94.4	97.2 ( $\uparrow$ 0.3 )	153.9	7.2	46.8 ( 1.15 $\times$ )
+ VLASH (d=4)	92.5	96.9	93.3	89.6	93.1 ( $\downarrow$ 3.8 )	176.7	5.8	32.8 ( 1.64 $\times$ )
+ StreamingVLA (d=1)	96.6	96.6	95.4	91.0	94.9 ( $\downarrow$ 2.0 )	n/a	n/a	31.6 ( 1.70 $\times$ )
+ FlashVLA (d=1)	98.8	99.6	97.6	95.4	97.8 ( $\uparrow$ 0.9 )	158.0	3.5	22.1 ( 2.43 $\times$ )

Table 1: Asynchronous performance on LIBERO at one-step delay (

d{=}1

) unless noted. FlashVLA reaches the largest speedup over

\pi_{0.5}

(2.43

\times

), while improving average success rate from 96.9% to 97.8%. Other asynchronous baselines either provide smaller speedups or lose success rate. Steps : actions executed per episode (averaged over 2,000 episodes); Time (s) : per-episode completion time.

Panel of Figure 3 — Figure 3: Success rate and time per step across asynchronous delay $d$ on LIBERO and RoboTwin 2.0. Both VLASH and FlashVLA are built on the same $\pi_{0.5}$ baseline. FlashVLA remains above the synchronous success rate across $d{=}1$ – $4$ while achieving lower time per step. Success rate is averaged over four LIBERO suites and the clean and randomized evaluations of 50 RoboTwin 2.0 tasks; dashed lines denote synchronous $\pi_{0.5}$ .

Method	LIBERO					RoboTwin 2.0
Method	Spatial	Object	Goal	Long	Avg.	Clean	Random	Avg.
$\pi_{0.5}$	98.8	98.2	98.0	92.4	96.9	86.1	85.8	86.0
+FlashVLA	98.6	99.0	97.8	96.2	97.9 ( $\uparrow$ 1.0)	90.8	90.2	90.5 ( $\uparrow$ 4.5)

Table 3: Synchronous success rates on LIBERO and RoboTwin 2.0. On LIBERO, FlashVLA improves

\pi_{0.5}

by 1.0 points on average, with a 3.8-point gain on LIBERO-Long. RoboTwin 2.0 reports 50-task multitask evaluation at

d{=}0

, improving the average across clean and randomized settings from 86.0% to 90.5%.

Method	Short Horizon			Medium Horizon			Long Horizon
Method	Clean	Random	Avg.	Clean	Random	Avg.	Clean	Random	Avg.
$\pi_{0.5}$	93.5	94.2	93.9	81.9	80.4	81.1	54.2	51.8	53.0
+FlashVLA	93.2	93.3	93.2 ( $\downarrow$ 0.7 )	86.0	84.8	85.4 ( $\uparrow$ 4.3 )	90.8	88.4	89.6 ( $\uparrow$ 36.6 )

Table 4: Synchronous performance on all 50 RoboTwin 2.0 tasks, grouped by task horizon. FlashVLA preserves short-horizon performance and improves average success by 4.3 points on medium-horizon tasks and 36.6 points on long-horizon tasks.

Method	Avg. SR (%)	Time / Step (ms)	Inference Latency (ms)
	LIBERO
SmolVLA	80.1	41.2	19.7
$+\,$ FlashVLA ( $d{=}0$ )	80.1 ( $\uparrow$ 0.0 )	29.7 ( 1.39 $\times$ )	10.1 ( 1.95 $\times$ )
$+\,$ FlashVLA ( $d{=}1$ )	79.5 ( $\downarrow$ 0.6 )	28.7 ( 1.44 $\times$ )	–
	RoboTwin 2.0
LingBot-VLA	85.2	56.7	70.6

Table 5: Cross-architecture generalization across SmolVLA and LingBot-VLA. With no or only minimal additional parameters, FlashVLA preserves SmolVLA success and improves LingBot-VLA by 3.4–4.1 points while reducing inference latency by 1.95

\times

–2.81

\times

. Here,

d{=}0

denotes synchronous execution and

d{=}1

denotes one-step asynchronous execution; delay-independent inference latency is reported only at

d{=}0

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Task	Pick and Place	Whiteboard Wiping	Table Cleaning
Instruction	Pick up the fruit and place it into the basket	Wipe all writings off the whiteboard	Clean the table by putting all objects into the basket

Appendix

Table 6: Real-world tasks and the natural-language instructions provided to the model.

Figure 6: Chunk-wise causal attention ablation on LIBERO. Removing the causal mask while keeping the same streaming buffer drops asynchronous success rate by roughly 10 points on average, isolating the chunk-wise causal mask as the active ingredient in FlashVLA’s asynchronous continuity.

Panel of Figure 7 — Figure 7: Chunk size $C$ ablation on LIBERO with buffer length $N{=}5$ fixed. (a) Success rate. (b) Per-step latency.

Panel of Figure 8 — Figure 8: Buffer length $N$ ablation on LIBERO with chunk size $C{=}10$ fixed. (a) Success rate. (b) Per-step latency.

Explore similar work

Mar 30, 2026

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

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

Yiran Shi, Dongqi Guo, Tianchen Zhao +8

Tsinghua University · Lenovo Group Ltd.

Vision-Language-Action Models Efficient VLA Model Inference

Ask PDF

Sep 28, 2026

RAVEL: Asynchronous Rolling Inference for Flow-Based Vision-Language-Action Models

Flow-based vision-language-action (VLA) models are highly effective for generalist robot manipulation, yet their reliance on computationally expensive VLM encoding and multi-step iterative action generation imposes a significant latency bottleneck. The resulting inference latency makes it difficult for robots to respond quickly, especially in dynamic environments. We address this limitation with RAVEL (Rolling Asynchronous VLA Enabling Low-Latency Control), an asynchronous inference framework that addresses the computational bottlenecks of both the VLM backbone and the action expert. To reduce the delay from multi-step action denoising, RAVEL allows near-term actions to be executed after a single denoising step by carrying partially denoised future actions forward in a rolling buffer. To avoid blocking on slow VLM encoding, RAVEL decouples VLM encoding from rolling action generation, allowing the action expert to operate continuously using the latest available VLM context, while a lightweight Fast Observation Pathway (FOP) directly conditions the action expert on current observations. Across simulated and real-world manipulation tasks, RAVEL consistently achieves substantially lower response latency while maintaining the task capability of the underlying VLA, enabling high-frequency and responsive closed-loop control.

Yuhan Chen, Ke Yu, Pengfei Liu +3

Zhejiang University · Nanyang Technological University

Efficient VLA Model Inference Vision-Language-Action Models VLMs for Robotics

Ask PDF

May 4, 2026

Understanding Asynchronous Inference Methods for Vision-Language-Action Models

Vision-Language-Action (VLA) models offer a promising path to generalist robot control, but their inference latency causes observation staleness when generated actions are executed asynchronously. Several methods have been proposed concurrently to mitigate this problem: inference-time inpainting (IT-RTC), training-time delay simulation (TT-RTC), future-state-aware conditioning (VLASH), and lightweight residual correction (A2C2). Each takes a fundamentally different approach, but they have so far been evaluated independently with different codebases, base policies, and protocols. We present a systematic comparison of these four methods under controlled conditions. We develop two unified codebases that integrate all methods with harmonized library and dataset versions, and we benchmark them on the Kinetix suite with MLPMixer policies and on the LIBERO manipulation benchmark with SmolVLA, sweeping inference delays up to

d=20

control steps. A2C2's per-step residual correction is the most effective method on Kinetix, holding above 90% solve rate up to

d=8

, and also leads on LIBERO from

d=4

onwards. IT-RTC is competitive at low delays but degrades sharply under long chunks (

H=30

) and high delays. TT-RTC is the most robust training-based method: stable across

d_\max

choices, generalizes beyond its training delay distribution, and adds zero inference overhead. VLASH exhibits a clear low-delay vs. high-delay trade-off governed by the fine-tuning delay range

[0,d_\max]

. Code is available at https://github.com/TheAyos/async-vla-inference

Ayoub Agouzoul

École polytechnique · Systems Group, ETH Zürich

Robotic Control Efficient VLM Inference Efficient VLA Model Inference Robotic Manipulation Vision-Language-Action Models

Ask PDF