Accelerating VLA Inference with Perturbation Informed Execution

Jehun Lee; Hajun Lee; Jimin Ryu; Youhan Lee; Beomseok Kang; Jonghyun Choi; Jae-Joon Kim

cs.ROOct 14, 2026

Accelerating VLA Inference with Perturbation Informed Execution

Authors: Jehun Lee, Hajun Lee, Jimin Ryu, Youhan Lee, Beomseok Kang, Jonghyun Choi, Jae-Joon Kim

Organizations: Seoul National University

Abstract

Action-chunking vision-language-action (VLA) models predict sequences of future robot actions. Under fixed-horizon execution, the robot executes a fixedlength prefix of each predicted chunk before invoking the policy again, leaving the remaining predictions unused. For a given number of executed actions, short execution horizons require more policy calls. Extending the execution horizon amortizes policy inference cost over more actions. However, uniformly applying long horizons across episodes can limit opportunities for precise corrections. To address this limitation, we propose PICES (Perturbation-Informed Chunk and Execution Horizon Selection), a training-free method that jointly selects an action chunk and the number of actions executed before the next policy call. Within a single policy call, the action expert generates a batch of action chunks from multiple initial noise samples and their locally perturbed counterparts, using the same conditioning features across all samples. PICES exploits differences between corresponding actions in each nominal chunk and its counterpart to determine candidate-specific execution horizons and select a nominal chunk for execution. Experiments on simulation benchmarks and real-world manipulation tasks demonstrate that PICES achieves speedup by up to 2.3x in wall-clock policy inference time on Jetson Orin and Thor relative to fixed horizon baselines while improving the success rate.

Figures & tables

Figure 1: High-level idea of the PICES. (a) Fixed execution horizon generates a nominal action chunk from a single initial noise sample at each policy call and executes a fixed-length actions before replanning. (b) PICES uses multiple initial-noise samples and their perturbations to determine execution horizon of nominal action chunk. Our method reduces the number of policy calls required to complete the same episode, from four to two in the illustrated example. (c) Success rate versus speedup comparison across methods, showing that PICES achieves wall-clock inference time speedup while preserving success rate.

Figure 4: Overview of PICES. Step 1: Draw $N$ initial noise samples and construct a perturbed variant of each sample. Step 2: The action expert processes the resulting $N(1+1)$ inputs in a single batch under the shared condition $C$ , producing $N$ nominal action chunks and their corresponding perturbed chunks. Step 3: Stepwise action sensitivity $G_{k}^{(n)}$ is computed by comparing each nominal chunk with its perturbed counterparts. Step 4: Compute a sensitivity score $S_{i}$ for each nominal chunk by averaging its sensitivity over the first $W$ steps and use this score to determine its execution horizon $h_{i}$ . PICES retains candidates whose scores are at or below the median and selects the nominal chunk with the longest execution horizon from the pool.

Figure 5: Open-loop evaluation of selected action horizons. Action RMSE is computed between each selected nominal action and the corresponding reference action, following the evaluation protocol of the motivational study. Each fixed-step point is based on the single action chunk. For $A^{3}$ , AAC, PICES, and the lowest- $S_{n}$ variant, the horizontal coordinate indicates the average of the horizons selected for individual method. PICES and the lowest- $S_{n}$ variant use $N=3$ nominal candidates; the latter selects the candidate with the smallest initial-window score ( $S_{n}$ ).

Method	Success rate (%) $\uparrow$							Policy calls $\downarrow$	Jetson Orin		Jetson Thor
Method	Pick Coke can	Pick object	Move near	Open drawer	Close drawer	Place in closed drawer	Average	Policy calls $\downarrow$	Single call (ms) $\downarrow$	Speedup $\uparrow$	Single call (ms) $\downarrow$	Speedup $\uparrow$
Fixed ( $h=1$ )	99	83	97	54	65	13	68.5	46.8	209.9	1.0	101.1	1.0
Fixed ( $h=2$ )	99	80	92	57	67	15	68.3	26.6		1.8		1.8
Fixed ( $h=4$ )	82	64	72	52	77	18	60.8	19.4		2.4		2.4
Fixed ( $h=6$ )	66	48	55	47	77	14	51.2	14.3		3.3		3.3
Fixed ( $h=8$ )	65	42	44	35	78	8	45.3	8.6		5.4		5.4

Table 1: Main results on SimplerEnv–Fractal with GR00T N1.7. We evaluate six Google Robot tasks with 100 episodes per task. We report task success rates (%) and mean policy calls over successful episodes, together with per-call GPU latency and inference-time speedup on Jetson Orin and Thor. Speedup is relative to default fixed execution horizon. The best success rates and speedups are highlighted in bold.

	$\pi_{0.5}$						GR00T N1.5
Method	Avg. SR (%) $\uparrow$	Avg. Policy Calls $\downarrow$	Jetson Orin		Jetson Thor		Avg. SR (%) $\uparrow$	Avg. Policy Calls $\downarrow$	Jetson Orin		Jetson Thor
			Single call	Speedup $\uparrow$	Single call	Speedup $\uparrow$			Single call	Speedup $\uparrow$	Single call	Speedup $\uparrow$
			(ms) $\downarrow$	Speedup $\uparrow$	(ms) $\downarrow$	Speedup $\uparrow$			(ms) $\downarrow$	Speedup $\uparrow$	(ms) $\downarrow$	Speedup $\uparrow$
Fixed ( $h=5$ )	96.5	30.2	297.2	1.0	126.7	1.0	92.9	30.1	149.3	1.0	62.8	1.0
AAC	96.4	20.6	359.2	1.2	152.4	1.2	93.3	24.2	282.5	0.7	107.2	0.7
AAC ∗	96.0	21.4	316.0	1.3	137.4	1.3	92.8	25.8	185.9	0.9	77.7	0.9

Table 2: Results on LIBERO with

\pi_{0.5}

and GR00T N1.5. We report average success rates (SR, %) and policy calls across the four LIBERO suites, together with per-call GPU latency and inference-time speedup on Jetson Orin and Thor.

Method	Success rate (%) $\uparrow$				Policy calls $\downarrow$				Jetson Orin		Jetson Thor
Method	Tape	Stapler	Cube	Average	Tape	Stapler	Cube	Average	Single call (ms) $\downarrow$	Speedup $\uparrow$	Single call (ms) $\downarrow$	Speedup $\uparrow$
Fixed ( $h=4$ )	97.2	77.8	58.3	77.8	614.8	236.8	485.9	445.8	298.5	0.39	126.7	0.39
Fixed ( $h=8$ )	86.1	94.4	91.7	90.7	222.9	104.8	196.9	174.9		1.00		1.00
Fixed ( $h=12$ )	88.9	80.6	88.9	86.1	120.8	74.4	111.1	102.1		1.71		1.71
Fixed ( $h=16$ )	77.8	66.7	86.1	76.9	95.8	68.6	74.8	76.4		2.29		2.29
PICES	88.9	94.4	97.2	93.5	148.9	72.6	134.5	118.7	314.7	1.39	137.3	1.36

Table 3: Real-world task results using UR10e arm with

\pi_{0.5}

. We report per-task and average success rates (SR, %) and policy calls for three tasks, together with per-call GPU latency and inference-time speedup on Jetson Orin and Thor. Inference-time speedup is computed from the product of average policy calls and per-call GPU latency on each device, relative to Fixed (

h=8

Method	Avg. SR (%) $\uparrow$	Avg. Policy Calls $\downarrow$	Jetson Orin		Jetson Thor
			Single call	Speedup $\uparrow$	Single call	Speedup $\uparrow$
			(ms) $\downarrow$	Speedup $\uparrow$	(ms) $\downarrow$	Speedup $\uparrow$
Fixed ( $h=1$ )	68.5	46.7	209.8	1.0	101.1	1.0
PICES , $N=1$	67.3	23.4	214.3	2.0	106.9	1.9
PICES , $N=3$	68.6	18.5	242.6	2.2	113.4	2.3
PICES , $N=5$	68.5	17.4	303.6	1.9	129.8	2.1

Table 4: Effect of the number of nominal candidates on SimplerEnv–Fractal with GR00T N1.7. We report average task success rates (%) and average policy calls, together with per-call GPU latency and inference-time speedup relative to fixed execution horizon on Jetson Orin and Thor.

Acceleration Method	Success rate (%) $\uparrow$					Policy calls $\downarrow$					Jetson Orin		Jetson Thor
Acceleration Method	Spatial	Goal	Object	Long	Avg.	Spatial	Goal	Object	Long	Avg.	Single call (ms) $\downarrow$	Speedup $\uparrow$	Single call (ms) $\downarrow$	Speedup $\uparrow$
Baseline, Fixed	99.0	97.4	97.2	92.2	96.5	20.6	21.8	27.7	50.6	30.2	297.2	1.0	126.7	1.0
PICES	99.0	98.4	99.6	92.2	97.3	13.9	14.4	18.0	33.4	19.9	311.4	1.5	131.6	1.5
FastV (40%)	99.0	97.4	98.0	89.6	96.0	20.6	27.7	22.3	50.1	30.2	237.5	1.3	106.3	1.2
FastV (40%) + PICES	98.4	99.2	97.0	92.2	96.7	14.0	18.0	14.7	33.5	20.1	247.6	1.8	111.8	1.7
FastV (50%)	97.6	98.8	95.2	90.6	95.6	21.2	27.6	22.3	51.2	30.6	217.1	1.4	97.3	1.3

Table 5: Compatibility of PICES with VLA acceleration methods on

\pi_{0.5}

. We report success rates (%) and average policy calls per episode across four LIBERO suites, together with per-call GPU latency and total inference-time speedup on Jetson Orin and Thor. Speedup is measured relative to Fixed (

h=5

) without visual token compression on the same device.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Jetson Orin					Jetson Thor
	$B=1$	$B=5$	$B=10$	$B=15$	$B=20$	$B=1$	$B=5$	$B=10$	$B=15$	$B=20$
$\pi_{0.5}$	310.4	361.6	464.1	581.1	676.6	127.2	149.8	179.3	215.3	267.0
GR00T N1.5	149.3	162.1	192.2	230.8	273.4	62.8	67.2	73.7	82.7	90.8
GR00T N1.7	209.9	235.4	301.6	361.6	401.1	101.1	111.5	124.5	148.3	166.2

Appendix

Table 6: GPU latency across batch sizes on Jetson Orin and Jetson Thor. We report the total VLA inference latency (ms) for batch sizes

B\in\{1,5,10,15,20\}

Model	Jetson Orin					Jetson Thor
	$B=1$	$B=5$	$B=10$	$B=15$	$B=20$	$B=1$	$B=5$	$B=10$	$B=15$	$B=20$
$\pi_{0.5}$	9.03	10.25	15.37	18.05	18.31	4.90	4.53	5.21	5.87	6.99
GR00T N1.5	1.08	2.56	4.29	5.43	6.20	0.96	0.87	1.08	1.32	1.40
GR00T N1.7	2.91	5.10	9.17	10.11	9.56	2.74	2.08	2.34	3.14	3.25

Appendix

Table 7: Additional latency per batch across batch sizes on Jetson Orin and Jetson Thor. We report per batch additional latency (ms) for batch sizes

B\in\{1,5,10,15,20\}

Setting	Jetson AGX Orin	Jetson AGX Thor
Jetson Linux	R36.5.0	R38.4.0
CUDA	12.6	13.0
PyTorch	2.8.0	2.13.0+cu130
TensorRT	10.3.0.30	10.13.3.9
Power mode	MAXN	MAXN
Clock settings	Fixed GPU: 1300.5 MHz Fixed EMC: 3199 MHz	Fixed GPU: 1575 MHz Fixed EMC: 4266 MHz

Appendix

Table 8: GPU environments and evaluation settings.

Figure 10: Real-world experimental setup and task objects. Left: The UR10e robot with wrist-mounted and third-person cameras. Right: A tape and stand, a stapler, red and green cubes used for Tape Hanging, Stapler Pressing, and Cube Stacking, respectively.

Figure 11: Task examples for real-world experiments

Hyperparameter / Setting	Value
Base checkpoint	$\pi_{0.5}$ -base (OpenPI)
# Training GPUs	4 $\times$ NVIDIA RTX PRO 6000 Blackwell
Total batch size	128 (32 per GPU)
# Train steps	5,000 steps ( $\approx$ 17.7 epochs)
Fine-tuning mode	Full fine-tuning (all parameters)
Action chunk horizon	16 steps

Appendix

Table 9: Training hyperparameters and real-world experiment settings for fine-tuning

\pi_{0.5}

on the UR10e platform.

Method	Success rate (%) $\uparrow$					Policy calls $\downarrow$					Jetson Orin		Jetson Thor
Method	Spatial	Goal	Object	Long	Average	Spatial	Goal	Object	Long	Average	Single call (ms) $\downarrow$	Speedup $\uparrow$	Single call (ms) $\downarrow$	Speedup $\uparrow$
$\pi_{0.5}$
Fixed ( $h=1$ )	99.0	92.8	91.4	84.4	91.9	102.8	110.6	145.7	272.9	158.0	297.2	0.2	126.7	0.2
Fixed ( $h=2$ )	99.0	95.4	94.8	86.6	94.0	51.2	55.3	71.7	132.1	77.6		0.4		0.4
Fixed ( $h=5$ )	98.0	97.6	98.2	92.0	96.5	20.6	21.8	27.7	50.6	30.2		1.0		1.0
Fixed ( $h=8$ )	99.4	98.6	98.2	93.4	97.4	13.1	13.7	17.2	31.5	18.9		1.6		1.6

Appendix

Table 10: LIBERO results with

\pi_{0.5}

and GR00T N1.5. We report per-suite and average success rates (%) and policy calls, together with single-call GPU latency and inference-time speedup on Jetson Orin and Thor. Fixed (

h=x

) executes

x

actions per policy call; single-call latency is shared across the five fixed-horizon settings for each model and device. Speedups are reported relative to Fixed (

h=5

Explore similar work

VLA Knows Its Limits: Adaptive Execution Horizons for Robot Policies
Feb 24, 2026Haoxuan Wang, Gengyu Zhang, Yan Yan +2Vision-Language-Action Models VLMs for Robotics
University of Illinois Chicago · Cisco Research
Understanding Asynchronous Inference Methods for Vision-Language-Action Models
May 4, 2026Ayoub AgouzoulRobotic Control Efficient VLM Inference
École polytechnique · Systems Group, ETH Zürich
When to Switch: Reliable Action-Chunk Extension for Vision-Language-Action Models
Oct 5, 2026Seonghoon Yu, Dongwon Kim, HyungRok Jung +4Efficient VLA Model Inference Robotic Manipulation
KAIST · GIST · POSTECH +1

Feb 24, 2026J/K move · Enter open · S save

VLA Knows Its Limits: Adaptive Execution Horizons for Robot Policies

Haoxuan Wang, Gengyu Zhang, Yan Yan, Ramana Rao Kompella+1

University of Illinois Chicago · Cisco Research

Vision-Language-Action Models VLMs for Robotics Test-Time Adaptation of VLA Models Action Chunking

Open paper AskPDF arXiv

Action chunking has recently emerged as a standard practice in flow-based Vision-Language-Action (VLA) models. However, the effect and choice of the execution horizon - the number of actions to be executed from each predicted chunk - remains underexplored. In this work, we first show that varying the execution horizon leads to substantial performance deviations, with performance initially improving and then declining as the horizon increases. To uncover the reasons, we analyze the cross- and self-attention weights in flow-based VLAs and reveal two key phenomena: (i) intra-chunk actions attend invariantly to vision-language tokens, limiting adaptability to environmental changes; and (ii) the initial and terminal action tokens serve as stable anchors, forming latent centers around which intermediate actions are organized. Motivated by these insights, we interpret action self-attention weights as a proxy for the model's predictive limit and propose AutoHorizon, the first test-time method that dynamically estimates the execution horizon for each predicted action chunk to adapt to changing perceptual conditions. Across simulated and real-world robotic manipulation tasks, AutoHorizon is performant, incurs negligible computational overhead, and generalizes across diverse tasks and flow-based models.

Abstract

Explore similar work

VLA Knows Its Limits: Adaptive Execution Horizons for Robot Policies

Understanding Asynchronous Inference Methods for Vision-Language-Action Models

When to Switch: Reliable Action-Chunk Extension for Vision-Language-Action Models