FASTER: Rethinking Real-Time Flow VLAs

Yuxiang Lu; Zhe Liu; Xianzhe Fan; Zhenya Yang; Jinghua Hou; Junyi Li; Kaixin Ding; Hengshuang Zhao

cs.ROMar 19, 2026

FASTER: Rethinking Real-Time Flow VLAs

Authors: Yuxiang Lu, Zhe Liu, Xianzhe Fan, Zhenya Yang, Jinghua Hou, Junyi Li, Kaixin Ding, Hengshuang Zhao

Organizations: The University of Hong Kong · ACE Robotics

Abstract

Real-time execution is crucial for deploying Vision-Language-Action (VLA) models in the physical world. Existing asynchronous inference methods primarily optimize trajectory smoothness, but neglect the critical latency in reacting to environmental changes. By rethinking the notion of reaction in action chunking policies, this paper presents a systematic analysis of the factors governing reaction time. We show that reaction time follows a uniform distribution determined jointly by the Time to First Action (TTFA) and the execution horizon. Moreover, we reveal that the standard practice of applying a constant schedule in flow-based VLAs can be inefficient and forces the system to complete all sampling steps before any movement can start, forming the bottleneck in reaction latency. To overcome this issue, we propose Fast Action Sampling for ImmediaTE Reaction (FASTER). By introducing a Horizon-Aware Schedule, FASTER adaptively prioritizes near-term actions during flow sampling, compressing the denoising of the immediate reaction by tenfold (e.g., in $π_{0.5}$ and X-VLA) into a single step, while preserving the quality of long-horizon trajectory. Coupled with a streaming client-server pipeline, FASTER substantially reduces the effective reaction latency on real robots, especially when deployed on consumer-grade GPUs. Real-world experiments, including a highly dynamic table tennis task, prove that FASTER unlocks substantially improved real-time responsiveness for generalist policies, enabling rapid generation of accurate and smooth trajectories.

Figures & tables

Figure 1: We propose FASTER to alleviate the reaction latency bottleneck in action chunking flow policies. By compressing the sampling iterations of the immediate reaction into a single step, FASTER (bottom) achieves a 10 $\times$ acceleration compared to original $\pi_{0.5}$ and X-VLA (top). This enables real-time responsiveness in highly dynamic real-world tasks such as playing table tennis. Moreover, FASTER is a plug-and-play solution for flow-based VLAs that integrates seamlessly into the standard fine-tuning pipeline and requires no architectural modifications.

Figure 2: Temporal pipelines of (a) synchronous and (b) asynchronous inference in a robotic system composed of an action chunking policy server and a robot client. As indicated by the best and worst cases, reaction time depends on both inference latency and the interval between consecutive inference-execution cycles. We also illustrate the decomposition of two adjacent action chunks to clarify the discretized inference delay

d

and the execution horizon

s

(bounded by

s_{\text{min}}

and

H-d

) in the asynchronous client.

Mode	Infer Interval	$\Delta t_{\text{react}}\sim D_{\text{react}}$	$\mathbb{E}[\Delta t_{\text{react}}]$	$s_{\text{min}}$
Sync	$\Delta t_{\text{infer}}+\Delta t_{\text{exec}}$	$\mathcal{U}(\Delta t_{\text{infer}},2*\Delta t_{\text{infer}}+\Delta t_{\text{exec}})$	$1.5\Delta t_{\text{infer}}+0.5\Delta t_{\text{exec}}$	–
Async	$\Delta t_{\text{exec}}$	$\mathcal{U}(\Delta t_{\text{infer}},\Delta t_{\text{infer}}+\Delta t_{\text{exec}})$	$\Delta t_{\text{infer}}+0.5*\Delta t_{\text{exec}}$	$\lceil\Delta t_{\text{infer}}/\Delta t_{\text{ctrl}}\rceil$

Table 1: Reaction characteristics of synchronous and asynchronous inference.

\mathcal{U}(a,b)

denotes a uniform distribution with lower and upper bounds

a

and

b

, so the expectation of

\Delta t_{\text{react}}

equals the midpoint

(a+b)/2

Panel of Figure 3 — Figure 3: Visualizations of (a) straightness $S(\mathbf{A})$ of the denoising path during sampling of the action chunk, and (b,c) differences between the intermediate clean action estimates $\tilde{\mathbf{A}}_{t}^{\tau\rightarrow 0}$ at each sampling timestep $\tau$ and the final output $\mathbf{A}_{t}^{0}$ , using the tasks “Pick Beverage” and “Fold Towel”, respectively.

Figure 4: Illustration of (a) constant timestep schedule used in conventional flow sampling and (b) Horizon-Aware Schedule (HAS) used in FASTER that allocates adaptive hit times across the action chunk and accelerates the sampling of early actions, enabling streaming output.

Model	Method	RTX 4090			RTX 4060
Model	Method	TTFA $\downarrow$	$s_{\min}$ $\downarrow$	$\mathbb{E}[\Delta t_{\text{react}}]$ $\downarrow$	TTFA $\downarrow$	$s_{\min}$ $\downarrow$	$\mathbb{E}[\Delta t_{\text{react}}]$ $\downarrow$
$\pi_{0.5}$	Sync	80.0 ±1.6 ms	3	170.0ms	303.3 ±0.8 ms	10	621.6ms
	Async	80.0 ±1.6 ms	3	130.0ms	303.3 ±0.8 ms	10	470.0ms
	FASTER	62.1 ±3.1 ms	3	112.1 ms	238.6 ±1.9 ms	8	371.9 ms
	Speedup	1.29 $\times$	–	1.16 $\times$	1.27 $\times$	1.25 $\times$	1.26 $\times$
X-VLA	Sync	113.7 ±0.8 ms	4	237.2ms	399.5 ±8.5 ms	12	799.2ms

Table 2: Comparison of reaction capability on RTX 4090 and RTX 4060 GPUs. “ Speedup ” denotes the relative gain of our method over the asynchronous baseline.

Model	Method	RTX 4090		RTX 4060
Model	Method	vs. Sync	vs. Async	vs. Sync	vs. Async
$\pi_{0.5}$	Async	0.72	-	0.74	-
$\pi_{0.5}$	FASTER	0.81	0.66	0.88	0.77
X-VLA	Async	0.73	-	0.75	-
X-VLA	FASTER	1.00	1.00	1.00	1.00

Table 3: Comparison of reaction speed from a probabilistic perspective on RTX 4090 and RTX 4060 GPUs. “ vs. Sync” denotes the probability that a given method has a shorter reaction time than Sync, and “ vs. Async” is defined analogously.

Figure 5: Comparison of real-world reaction speed on the table tennis task. Left : Visualization of rollouts on RTX 4090, the third column corresponds to the contact moment, and the interval between each image in a row is 166.7ms (5 frames). Right : Quantitative completion scores on two GPUs.

Figure 6: Comparison of real-world performance and completion duration on two additional tasks.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Temporal pipeline of asynchronous inference at a fine-grained level. Suppose the inference latency

\Delta t_{\text{infer}}

is 2.5 times the controller period

\Delta t_{\text{ctrl}}

, resulting in an inference delay of

d=2

and a minimal execution horizon of

s_{\text{min}}=3

Panel of Figure 8 — Figure 8: Additional visualizations of (a,c,e) straightness $S(\mathbf{A})$ and (b,d,f) differences between the intermediate clean action estimates and the final output. Panels (a,b) are computed using a $\pi_{0.5}$ model fine-tuned on Pick Beverage task with prediction horizon $H=30$ ; (c,d) use an X-VLA model fine-tuned on the LIBERO simulation dataset with $H=30$ ; and (e,f) use an X-VLA model fine-tuned on the CALVIN ABC simulation dataset with $H=30$ . The shaded regions in (a,c,e) denote the $\pm 2$ SEM across 200 random samples.

GPU	$\pi_{0.5}$			X-VLA
GPU	Index	Time Req.	Time Rec.	Index	Time Req.	Time Rec.
RTX 4090	1	66.7ms	62.1 ±3.1 ms	1	66.7ms	44.8 ±0.3 ms
	2	100.0ms	70.2 ±3.1 ms	2	100.0ms	52.0 ±0.1 ms
	3	133.3ms	77.1 ±3.2 ms	3	133.3ms	59.6 ±0.3 ms
RTX 4060	7	266.7ms	238.6 ±1.9 ms	3	133.3ms	129.2 ±2.4 ms
	8	300.0ms	253.3 ±2.5 ms	4	166.7ms	159.0 ±3.6 ms

Appendix

Table 4: Illustration of inference latency vs. sequential execution. “Index” denotes the index of valid actions in the chunk, excluding delayed actions. “Time Req.” is the time at which an action is required by the robot controller, while “Time Rec.” is the time at which an action is received from the policy server via the streaming interface, reported as mean

\pm

standard deviation over 20 trials. All timestamps are measured relative to the start of inference.

\mathbf{m}=\{m_{i}\}=\mathbf{1}(i\geq d),\quad i\in[0,H-1],

Appendix

Algorithm 1 FASTER policy training

\mathbf{u}=\{u_{i}\}=\left(1-\left(\frac{i-d}{\max(H-1-d,1)}\right)^{\alpha}\right)\cdot u_{d},\quad i\in[d+1,H-1],

Appendix

Algorithm 2 FASTER policy inference

Figure 9: AgileX Cobot Magic robotic platform with Piper arms.

Figure 10: Visualization of Pick Beverage and Fold Towel tasks.

Model	$\pi_{0.5}$		X-VLA
Task	AgileX	Simulation	AgileX	Simulation
Prediction Horizon $H$	50	10	30	30
Action Space	Relative Joint	Delta EEF	Abs EE6D	Abs EE6D
Global Batch Size	128	256	128	128
Training Steps	50k	30k	50k	60k
Optimizer	AdamW	AdamW	AdamW	AdamW

Appendix

Table 5: Hyperparameter settings for fine-tuning VLA models.

Task	Method	RTX 4090		RTX 4060
Task	Method	$d$	$s$	$d$	$s$
Table Tennis	Sync	4	5	10	11
	Async	4	5	10	11
	FASTER	3	4	8	10
Pick Beverage, Fold Towel	Sync	4	50	10	50
	Async	4	46	10	40

Appendix

Table 6: Settings of inference delay

d

and execution horizon

s

on RTX 4090 and RTX 4060 GPUs. “Async” refers to Naive Async and Training-time RTC.

Model	Method	RTX 4090			RTX 4060
Model	Method	$\Delta t_{\text{infer}}$	$\Delta t_{\text{exec}}$	$D_{\text{react}}$	$\Delta t_{\text{infer}}$	$\Delta t_{\text{exec}}$	$D_{\text{react}}$
$\pi_{0.5}$	Sync	80.0ms	100.0ms	$\mathcal{U}(80.0,260.0)$	303.3ms	333.3ms	$\mathcal{U}(303.3,939.9)$
	Async	80.0ms	100.0ms	$\mathcal{U}(80.0,180.0)$	303.3ms	333.3ms	$\mathcal{U}(303.3,636.6)$
	FASTER	62.1 ms	100.0ms	$\mathcal{U}(62.1,162.1)$	238.6 ms	266.7 ms	$\mathcal{U}(238.6,505.3)$
X-VLA	Sync	113.7ms	133.3ms	$\mathcal{U}(113.7,360.7)$	399.5ms	400.0ms	$\mathcal{U}(399.5,1199.0)$
	Async	113.7ms	133.3ms	$\mathcal{U}(113.7,247.0)$	399.5ms	400.0ms	$\mathcal{U}(399.5,799.5)$

Appendix

Table 7: Detailed comparison of reaction capability on RTX 4090 and RTX 4060 GPUs. The inference latency

\Delta t_{\text{infer}}

corresponds to TTFA, and execution duration

\Delta t_{\text{exec}}

is computed as

s_{\text{min}}\cdot\Delta t_{\text{ctrl}}

. The distribution of reaction time

D_{\text{react}}

\mathcal{U}(\Delta t_{\text{infer}},2*\Delta t_{\text{infer}}+\Delta t_{\text{exec}})

for Sync, and

\mathcal{U}(\Delta t_{\text{infer}},\Delta t_{\text{infer}}+\Delta t_{\text{exec}})

for Async and FASTER.

Figure 11: Comparison of real-world performance and task completion duration on two real-world tasks using X-VLA. Note that the duration is computed only from successful rollouts, and therefore is not directly comparable to the results in Figure 6 .

Model	Method	RTX 4090		RTX 4060
Model	Method	Mean	96% CI	Mean	96% CI
$\pi_{0.5}$	Sync	0.000	(0.000, 0.000)	0.000	(0.000, 0.000)
	Naive Async	0.200	(0.075, 0.325)	0.200	(0.067, 0.333)
	Training-time RTC	0.533	(0.300, 0.767)	0.300	(0.133, 0.500)
	FASTER	0.800	(0.667, 0.933)	0.467	(0.333, 0.567)

Appendix

Table 8: Performance statistics for the Table Tennis task. The 96% confidence intervals are computed using bootstrap resampling.

Model	Method	Pick Beverage				Fold Towel
		Score		Duration		Score		Duration
		Mean	96% CI	Mean	Std	Mean	96% CI	Mean	Std
$\pi_{0.5}$	Sync	0.879	(0.786, 0.950)	13.0	0.9	0.788	(0.600, 0.925)	24.7	0.5
	Naive Async	0.957	(0.886, 1.000)	12.5	1.2	0.825	(0.613, 0.988)	24.0	2.4
	Training-time RTC	0.950	(0.879, 0.993)	11.9	2.4	0.888	(0.700, 1.000)	20.7	0.4
	FASTER	0.957	(0.886, 1.000)	12.0	0.9	0.963	(0.925, 1.000)	20.5	0.4

Appendix

Table 9: Performance statistics for the Pick Beverage and Fold Towel task. The 96% confidence intervals are computed using bootstrap resampling.

Method	LIBERO					CALVIN ABC $\rightarrow$ D
Method	Spatial	Object	Goal	10	Avg.	1	2	3	4	5	Avg. Len
$\pi_{0.5}$	98.8	98.2	98.0	92.4	96.9	94.2	88.7	85.7	83.2	79.5	4.313
$\pi_{0.5}$ +FASTER	98.6	97.8	97.8	91.6	96.5	95.1	89.1	85.0	81.9	78.1	4.292
X-VLA ∗	97.8	99.4	97.8	96.8	98.0	95.7	89.8	82.4	77.0	70.2	4.151
X-VLA+FASTER	97.8	99.4	98.2	93.0	97.1	98.1	91.4	82.4	72.9	64.5	4.093

Appendix

Table 10: Comparison of performance on simulation benchmarks LIBERO and CALVIN. “

*

” indicates experimental results evaluated on our cluster with official checkpoints.

Method	$d$	$s$	Solve Rate
Naive Async	4	4	0.492
BID [ 57 ]	4	4	0.553
Inference-time RTC [ 4 ]	4	4	0.614
Training-time RTC [ 5 ]	4	4	0.726
REMAC [ 93 ]	4	4	0.779
VLASH [ 88 ]	4	4	0.813

Appendix

Table 11: Comparison of performance on Kinetix benchmark under the same inference budget.

$p$	CALVIN ABC $\rightarrow$ D
$p$	1	2	3	4	5	Avg. Len
Baseline	95.7	89.8	82.4	77.0	70.2	4.151
0.3	93.7	85.2	76.0	65.5	55.2	3.756
0.5	98.1	91.4	82.4	72.9	64.5	4.093
0.7	89.6	76.7	63.2	50.8	40.3	3.206
1.0 (w/o mixed)	89.0	74.7	60.4	49.0	38.1	3.112

Appendix

Table 13: Ablation study of mixing probability

p

in the mixed schedule, using X-VLA on the CALVIN benchmark. Factor

\alpha

is set to 0.7.

Figure 13: Open-loop action prediction error on the Fold Towel task. We compare the $\pi_{0.5}$ baseline with FASTER. The heatmaps report the mean absolute error for each action dimension and action chunk index, normalized by the maximum error of each action dimension computed over 200 random samples and shared across both models.

Efficient VLA Model Inference

Efficient vision-language-action (VLA) model inference reduces robot-control latency, memory use, and energy while preserving task performance.

See all 190 papers

Jun 12, 2026

ReactVLA: Fast and Lightweight Reactive Robot Manipulation via Improved Mean Flow Action Generation

Diffusion-based Vision-Language-Action (VLA) policies have demonstrated strong capability in modeling expressive and multimodal action distributions. However, their reliance on iterative sampling introduces substantial inference latency, which limits their applicability to reactive closed-loop robot manipulation. To address this limitation, we propose ReactVLA, a lightweight and low-latency VLA framework for real-time robotic manipulation. ReactVLA combines two complementary designs: (1) an improved Mean Flow (iMF) action generator that reduces expensive multi-step diffusion sampling to one-to-few-step action generation, and (2) Attention Residuals (AttnRes), a dynamic depth-wise feature routing mechanism that replaces uniform residual accumulation to better preserve task-relevant multimodal representations. We evaluate ReactVLA on large-scale simulation benchmarks, including LIBERO and RoboIMI, as well as real-world robotic manipulation tasks. Experimental results show that ReactVLA consistently outperforms similarly sized VLA baselines, including SmolVLA and

π_0

. On challenging precision manipulation tasks, ReactVLA achieves up to a 1.65

\times

improvement in task performance while providing more than a 4

\times

increase in inference speed compared with leading VLA models. Finally, it reduces real-world policy latency to below 38.6 ms, enabling fast reactive control on physical robot platforms. Please check out our project website at: https://game-loader.github.io/ReactVLA/.

Yanzhao Guo, Wenkai Chen, Jianwei Zhang

Shanghai Jiao Tong University · Technical Aspects of Multimodal Systems (TAMS), Department of Informatics, Universität Hamburg

Language-Conditioned Robot Manipulation Efficient VLM Inference Attention Residual Connections Vision-Language-Action Models

Ask PDF

Aug 27, 2026

FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference

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.

Zekai Li, Jiaming Tang, Zhijian Liu

UC San Diego · MIT

Efficient VLA Model Inference VLMs for Robotics Action Chunking

Ask PDF

Jul 14, 2026

Reducing Temporal Redundancy for Efficient Vision-Language-Action Inference

Vision-Language-Action (VLA) models exhibit strong generalization for robotic manipulation, yet their high inference latency limits real time deployment. We identify two primary sources of temporal redundancy in existing VLA pipelines: repeated visual encoding of highly similar consecutive frames and multi step iterative sampling in diffusion based policies. To address this, we propose a system level acceleration strategy that reduces computation in both perception and action generation. On the perception side, we incrementally update only tokens corresponding to dynamic scene regions instead of re-encoding entire frames. On the policy side, we compress diffusion sampling into a compact 2-step schedule through efficiency oriented training while preserving action precision. Experiments on Libero, RobotWin, and Real Robot Platforms demonstrate over 2 times speedup while maintaining high performance, achieving up to 98% success rate on general manipulation benchmarks. Our codes will be released on Github.

Yuzhou Wu, Yuxin Zheng, Muchun Niu +6

1Tianji KernalMind co ltd · 2Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · 3Shanghai Jiao Tong University, Shanghai, China +2

Efficient VLA Model Inference Efficient VLA Models Robotic Manipulation Vision-Language-Action Models Diffusion Model Inference Acceleration

Ask PDF

Abstract

Efficient VLA Model Inference

Explore similar work

ReactVLA: Fast and Lightweight Reactive Robot Manipulation via Improved Mean Flow Action Generation

FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference

Reducing Temporal Redundancy for Efficient Vision-Language-Action Inference