DepthCache: Depth-Guided Training-Free Visual Token Merging for Vision-Language-Action Model Inference

Yuquan Li; Lianjie Ma; Han Ding; Lijun Zhu

cs.ROMar 11, 2026

DepthCache: Depth-Guided Training-Free Visual Token Merging for Vision-Language-Action Model Inference

Authors: Yuquan Li, Lianjie Ma, Han Ding, Lijun Zhu

Organizations: School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China. · School of Mechanical Science and Engineering, Huazhong University of Science and Technology, Wuhan, China.

Abstract

Vision-Language-Action (VLA) models enable generalist robotic manipulation but suffer from high inference latency. This bottleneck stems from the massive number of visual tokens processed by large language backbones. Existing methods either prune or merge tokens uniformly, degrading the spatial reasoning essential for robotic control. We present DepthCache, a training-free framework that leverages depth as a structural prior for visual token compression. It partitions observations into depth-based regions and applies spatially differentiated merge ratios, preserving the near-field workspace while compressing the distant background. To exploit temporal redundancy, DepthCache distributes the merging process across consecutive frames, ensuring consistent representations while reducing per-step computation. A motion-adaptive pipeline further optimizes auxiliary view compression based on end-effector dynamics. The framework requires no model modification, generalizing across diverse VLA architectures. On the LIBERO benchmark, DepthCache achieves up to 1.28x inference speedup with less than 1% average success rate degradation across three VLA models (pi_0.5, OpenVLA, GR00T), whereas pruning and merging baselines incur 4--24% degradation at comparable compression. Real-world experiments on a physical manipulator demonstrate that DepthCache enables faster task throughput and more responsive closed-loop control in latency-sensitive scenarios.

Figures & tables

Fig. 1 : DepthCache enables training-free inference acceleration for VLA models with negligible performance loss. Top: speedup and success rate on LIBERO across three VLA architectures. Bottom: directly transplanting token merging (ToSA, left) causes hesitation during grasping; DepthCache (right) maintains fluid execution with reduced latency.

Fig. 2 : Overview of DepthCache. The primary view pipeline (top) forms a cyclic process: initialization computes dual protection sets, depth-based partitioning assigns differentiated merge ratios, and progressive merging reduces tokens across frames until scene change triggers re-initialization. The auxiliary view pipeline (bottom) employs a two-state machine to gate wrist camera compression by end-effector motion.

	Success Rate (%)					Efficiency
Method	Spatial	Object	Goal	Long	Avg.	Speedup	$\rho$
OpenVLA	83.3	91.2	78.6	53.7	76.7	1.00 $\times$	100
OpenVLA + FastV	75.8	70.3	67.3	42.5	64.0 -12.7	1.39 $\times$	75.0
OpenVLA + SP-VLA †	80.0	82.4	73.6	51.6	71.9 -4.8	1.50 $\times$	dyn. ‡
OpenVLA + DepthCache	83.1	90.7	77.8	51.2	75.7 -1.0	1.21 $\times$	78.9
$\pi_{0.5}$	99.0	97.0	99.0	96.5	97.9	1.00 $\times$	100

TABLE I : Main results on LIBERO simulation benchmark.

Task	Successes / 20
Task	Baseline	+ Ours
Pick & Place	20 / 20	20 / 20
Stack Blocks	18 / 20	17 / 20
Drawer & Place	17 / 20	15 / 20
Total	55 / 60	52 / 60
Avg. Latency (ms)	191	143

TABLE II : Real-world core task results with

\pi_{0.5}

Panel of Figure 5 — Fig. 3 : Real-world core task execution with DepthCache-augmented $\pi_{0.5}$ . Perturbation recovery is shown in Fig. 1 .

Fig. 4 : Sequential multi-object sorting comparison. Top: baseline $\pi_{0.5}$ . Bottom: $\pi_{0.5}$ + DepthCache. The time gap accumulates across successive pick-and-place cycles.

Scenario	Successes / 15		Time (s)
Scenario	Baseline	+ Ours	Baseline	+ Ours
Multi-Object Sorting	15 / 15	13 / 15	28.6	22.1
Perturbation Recovery	11 / 15	12 / 15	17.4	13.7

TABLE III : Extended scenario results (15 trials each).

Fig. 5 : Parameter sensitivity analysis on

\pi_{0.5}

(LIBERO). (a) Effect of maximum merge ratio

r_{\max}

on success rate and speedup. (b) Effect of progressive merge step ratio

\eta

on success rate (log-scale

x

-axis). Default values (

r_{\max}=0.7

\eta=0.2

) are highlighted.

Configuration	Success Rate (%)						Efficiency
Configuration	Spatial	Object	Goal	Long	Avg.	$\Delta$	Speedup	$\rho$
Full DepthCache	98.9	97.1	98.3	96.1	97.6	—	1.28 $\times$	68.2
w/o Depth Partitioning	92.9	75.1	78.4	71.1	79.4	-18.2	1.34 $\times$	67.5
w/o Dual Protection	96.0	85.7	92.1	87.2	90.3	-7.3	1.48 $\times$	55.6
w/o Progressive Merge	94.7	68.1	82.5	78.6	81.0	-16.6	1.32 $\times$	65.4
w/o Re-initialization	91.8	97.2	92.6	89.4	92.8	-4.8	1.25 $\times$	72.5

TABLE IV : Ablation study on DepthCache components using

\pi_{0.5}

Explore similar work

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
Text-Vision Synergistic Token Caching: A Training-Free Framework for Efficient Vision-Language-Action Inference
Sep 28, 2026Qianer Li, Chengjie Zhang, Jingwen Chen +3Efficient VLA Model Inference Vision-Language-Action Models
College of Engineering, Southern University of Science and Technology (SUSTech), Shenzhen 518055, China · Master of Science in Data Science and Machine Learning, Department of Mathematics, Faculty of Science, National University of Singapore, Singapore
Training-Free Acceleration for Vision-Language-Action Models with Action Caching and Refinement
Jul 7, 2026Ryuji Oi, Hikari Otsuka, Kosuke Matsushima +4Efficient VLA Model Inference Vision-Language-Action Models
Institute of Science Tokyo

Jul 14, 2026J/K move · Enter open · S save

Reducing Temporal Redundancy for Efficient Vision-Language-Action Inference

Yuzhou Wu, Yuxin Zheng, Muchun Niu, Yishan Yang+5

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

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

Reducing Temporal Redundancy for Efficient Vision-Language-Action Inference

Text-Vision Synergistic Token Caching: A Training-Free Framework for Efficient Vision-Language-Action Inference

Training-Free Acceleration for Vision-Language-Action Models with Action Caching and Refinement