DySL-VLA: Efficient Vision-Language-Action Model Inference via Dynamic-Static Layer-Skipping for Robot Manipulation

Zebin Yang; Yijiahao Qi; Tong Xie; Bo Yu; Shaoshan Liu; Meng Li

cs.ROFeb 26, 2026

DySL-VLA: Efficient Vision-Language-Action Model Inference via Dynamic-Static Layer-Skipping for Robot Manipulation

Authors: Zebin Yang, Yijiahao Qi, Tong Xie, Bo Yu, Shaoshan Liu, Meng Li

Organizations: Institute for Artificial Intelligence, Peking University, Beijing, China · School of Integrated Circuits, Peking University, Beijing, China · School of Electronics Engineering and Computer Science, Peking University, Beijing, China · Shenzhen Institute of Artificial Intelligence and Robotics for Society, Shenzhen, China · Beijing Advanced Innovation Center for Integrated Circuits, Beijing, China

Abstract

Vision-Language-Action (VLA) models have shown remarkable success in robotic tasks like manipulation by fusing a language model's reasoning with a vision model's 3D understanding. However, their high computational cost remains a major obstacle for real-world applications that require real-time performance. We observe that the actions within a task have varying levels of importance: critical steps demand high precision, while less important ones can tolerate more variance. Leveraging this insight, we propose DySL-VLA, a novel framework that addresses computational cost by dynamically skipping VLA layers based on each action's importance. DySL-VLA categorizes its layers into two types: informative layers, which are consistently executed, and incremental layers, which can be selectively skipped. To intelligently skip layers without sacrificing accuracy, we invent a prior-post skipping guidance mechanism to determine when to initiate layer-skipping. We also propose a skip-aware two-stage knowledge distillation algorithm to efficiently train a standard VLA into a DySL-VLA. Our experiments indicate that DySL-VLA achieves 2.1% improvement in success length over Deer-VLA on the Calvin dataset, while simultaneously reducing trainable parameters by a factor of 85.7 and providing a 3.75x speedup relative to the RoboFlamingo baseline at iso-accuracy. Our code is available on https://github.com/PKU-SEC-Lab/DYSL_VLA.

Figures & tables

Figure 1. Different actions in robot manipulation have different importance. We show an example when the robot is performing task “Grasp the black cup and drop it into basket”. (a) shows the task completion rates when adding noise with different magnitudes to VLA model weights at different action steps. When adding noise at important action steps, the task completion rate drops faster as noise magnitude increases. We sample 50 times on each noise magnitude for each step range. We show the robot status at (b) step 25, (c) step 75, and (d) step 125 when using the origin VLA model.

Method	Approach	Action Importance	Specialized	Training
Method	Approach	Aware	Kernel Free	Modules
SparseVLM ( Zhang et al., 2024b )	Pruning	✗	✗	–
FastV ( Chen et al., 2024 )	Pruning	✗	✗	–
QAIL ( Park et al., 2024 )	Quantization	✗	✗	LLM Backbone
MoLe-VLA ( Zhang et al., 2025 )	Mixture-of-Layers	✗	✓	LLM Backbone
DeeR-VLA ( Yue et al., 2024 )	Early-exit	✓	✓	LLM Backbone

Table 1. Comparison between different methods.

Figure 3. The average cosine similarity between the output activations of different VLA layers for (a) RoboFlamingo-3B and (b) RoboFlamingo-9B. The similarity between the input and output activations of each layer and the model performance when skipping each VLA layer in a zero-shot manner for (c) RoboFlamingo-3B and (d) RoboFlamingo-9B.

Figure 4. (a) The ratio of different numbers of kept layers in VLA model inference when only using skipping controllers. The inference latency for different numbers of kept layers using (b) RoboFlamingo-3B in FP32 and (c) RoboFlamingo-9B in FP16.

Figure 5. The inference mode of (a) original VLA model, (b) using early exit with multiple action heads, (c) using early exit with adapters, (d) using traditional layer skipping methods, and (e) using dynamic-static layer skipping. The modules with light colour are not activated in current inference. We set the same legend for VLA layers in (a), (b), (c), (d), and dynamic layers in (e).

Figure 6. The proportion of action prediction steps in different continuity ranges and the accuracy loss when conducting layer skipping at the steps in different continuity ranges for (a) RoboFlamingo-3B and (b) RoboFlamingo-9B.

Figure 7. Prior-post skipping guidance. The modules with light colour are not activated in current inference.

Figure 8. Skipping-allow point changes in different steps. Here we use 4 dynamic layers between 2 static layers as an example, with 5 possible positions for skipping-allow point.

Figure 9. The (a) first stage and (b) second stage of skip-aware two-stage knowledge distillation method. The layers in red boxes are the selected layers in current training step.

Method	# Fine-tuned	# Fine-tuned	Training Cost	D $\rightarrow$ D	ABC $\rightarrow$ D	RTX 4090
Method	Parameters	Steps	(GPU Hour)	Avg Length	Avg Length	Latency
HULC ( Mees et al., 2022a )	–	–	–	2.64	0.67	–
SPIL ( Zhou et al., 2024 )	–	–	–	2.67	1.71	–
RoboFlamingo 3B ( Li et al., 2023 )	–	–	–	2.92	2.85	51.0ms
DeeR-VLA 3B ( Yue et al., 2024 )	1.2B	$9.2\cdot 10^{4}$	112	2.83	2.82	19.3ms
Random Skip	–	–	–	0.38	0.45	22.6ms

Table 2. Evaluation on Calvin dataset.

Method	# Fine-tuned	Spatial	Object	Goal	Long	Average	A6000	Jetson Orin
Method	Parameters	SR (%)	SR (%)	SR (%)	SR (%)	SR (%)	Latency	Latency
MDT (scratch) ( Reuss et al., 2024 )	–	78.5	87.5	73.5	64.8	76.1	–	–
$\pi_{0}$ + FAST ( Black et al., 2024 )	–	96.4	96.8	88.6	60.2	85.5	–	–
OpenVLA-OFT 7B ( Kim et al., 2025 )	–	97.6	98.4	97.9	94.5	97.1	53.0ms	676ms
DeeR-VLA 7B ( Yue et al., 2024 )	7.1B	97.0	98.2	97.4	88.6	95.3	40.2ms	495ms
Random Skip	–	11.3	0.6	0.0	0.6	3.1	34.9ms	383ms

Table 3. Evaluation on LIBERO dataset.

Method	Avg Length	Avg Latency
RoboFlamingo 3B ( Li et al., 2023 )	2.82	51.0ms
Random Skip	0.38	22.6ms
DySL-VLA	2.89	13.6ms
w/o Post-skip Verification	2.79	13.4ms
w/o Pre-skip Prediction	2.42	20.1ms
w/o Skip-aware Two-stage Knowledge Distillation	2.82	74.0ms

Table 4. Individual influence of our methods (evaluated on Calvin D

\rightarrow

D dataset).

Figure 10. Latency breakdown of LLM backbone.

Static Layer Ratio	10%	15%	20%	25%	30%
Average Length	2.81	2.84	2.89	2.88	2.89
Average Latency (ms)	12.6	13.1	13.6	14.7	15.8

Table 5. Ablation study on static layer ratio, evaluated on Calvin D

\rightarrow

D dataset.

$\delta l$	1	2	3	4	5	Adaptive
SR	96.2	96.4	96.4	97.4	97.0	98.0

Table 6. Ablation study on

\delta l

(using LIBERO Spatial dataset).

$k$	1	3	5	7	9
Average Length	2.59	2.76	2.89	2.80	2.69

Table 7. Influence of the value of

k

in pre-skip prediction, evaluated on Calvin D

\rightarrow

D dataset.

Explore similar work

A1: A Fully Transparent Open-Source, Adaptive and Efficient Truncated Vision-Language-Action Model
Apr 7, 2026Kaidong Zhang, Jian Zhang, Rongtao Xu +20Vision-Language-Action Models Efficient VLA Model Inference
SYSU · Spatialtemporal AI · MBZUAI
Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think
Jun 18, 2026Gia-Binh Nguyen, Trong-Bao Ho, Thien-Loc Ha +18Model Compression Efficient VLA Model Inference
Center for AI Research, VinUniversity · VinRobotics · University of Arkansas +10
Reducing Temporal Redundancy for Efficient Vision-Language-Action Inference
Jul 14, 2026Yuzhou Wu, Yuxin Zheng, Muchun Niu +6Efficient VLA Model Inference Efficient VLA Models
1Tianji KernalMind co ltd · 2Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · 3Shanghai Jiao Tong University, Shanghai, China +2

Apr 7, 2026J/K move · Enter open · S save

A1: A Fully Transparent Open-Source, Adaptive and Efficient Truncated Vision-Language-Action Model

Kaidong Zhang, Jian Zhang, Rongtao Xu, Yu Sun+19

SYSU · Spatialtemporal AI · MBZUAI

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

Open paper AskPDF arXiv

Vision-Language-Action (VLA) models have emerged as a powerful paradigm for open-world robot manipulation, but their practical deployment is often constrained by cost: billion-scale VLM backbones and iterative diffusion/flow-based action heads incur high latency and compute, making real-time control expensive on commodity hardware. We present A1, a fully open-source and transparent VLA framework designed for low-cost, high-throughput inference without sacrificing manipulation success; Our approach leverages pretrained VLMs that provide implicit affordance priors for action generation. We release the full training stack (training code, data/data-processing pipeline, intermediate checkpoints, and evaluation scripts) to enable end-to-end reproducibility. Beyond optimizing the VLM alone, A1 targets the full inference pipeline by introducing a budget-aware adaptive inference scheme that jointly accelerates the backbone and the action head. Specifically, we monitor action consistency across intermediate VLM layers to trigger early termination, and propose Inter-Layer Truncated Flow Matching that warm-starts denoising across layers, enabling accurate actions with substantially fewer effective denoising iterations. Across simulation benchmarks (LIBERO, VLABench) and real robots (Franka, AgiBot), A1 achieves state-of-the-art success rates while significantly reducing inference cost (e.g., up to 72% lower per-episode latency for flow-matching inference and up to 76.6% backbone computation reduction with minor performance degradation). On RoboChallenge, A1 achieves an average success rate of 29.00%, outperforming baselines including pi0(28.33%), X-VLA (21.33%), and RDT-1B (15.00%).

Abstract

Explore similar work

A1: A Fully Transparent Open-Source, Adaptive and Efficient Truncated Vision-Language-Action Model

Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think

Reducing Temporal Redundancy for Efficient Vision-Language-Action Inference