RAPID: Redundancy-Aware and Compatibility-Optimal Edge-Cloud Partitioned Inference for Diverse VLA Models

Zihao Zheng; Sicheng Tian; Hangyu Cao; Chenyue Li; Jiayu Chen; Maoliang Li; Xinhao Sun; Hailong Zou

cs.DCMar 9, 2026

RAPID: Redundancy-Aware and Compatibility-Optimal Edge-Cloud Partitioned Inference for Diverse VLA Models

Authors: Zihao Zheng, Sicheng Tian, Hangyu Cao, Chenyue Li, Jiayu Chen, Maoliang Li, Xinhao Sun, Hailong Zou,

Organizations: School of Computer Science, Peking University, Beijing, China. · School of Artificial Intelligence, Beijing Normal University, Beijing, China. · School of Software Engineering, South China University of Technology, Guangzhou, China. · School of Electronics Engineering and Computer Science, Peking University, Beijing China.

Abstract

Vision Language Action (VLA) models are mainstream in embodied intelligence but face high inference costs. Edge-Cloud Collaborative (ECC) inference offers an effective fix by easing edge-device computing pressure to meet real-time needs. However, existing ECC frameworks are suboptimal for VLA models due to two challenges: (1) Mainstream environment-oriented edge-cloud partitioning methods are susceptible to interference from visual noise; (2) Existing edge-cloud partitioning methods overlook the step-wise redundancy unique to embodied tasks, thereby disrupting the physical continuity of motion. To address these issues, we propose a novel ECC inference framework, termed RAPID. Specifically, we developed an implementation tailored to the proposed framework. Experiments demonstrate this achieves a speedup of up to 1.73x with only 5%~7% overhead.

Figures & tables

Fig. 1: Comparison between Vision-Based Strategy(Left) and Our RAPID Framework(Middle).

Noise	Cloud-Side		Edge-Side		Total
Noise	Lat.	Load	Lat.	Load	Lat.	Load
Standard	62.5ms	9.5GB	315.2ms	4.7GB	395.4ms	14.2GB
Visual Noise	75.4ms	11.2GB	210.5ms	3.0GB	520.6ms	14.2GB
Distract.	88.6ms	13.0GB	95.4ms	1.2GB	685.3ms	14.2GB

TABLE I: Performance of Vision-Based Dynamic Strategy under Different Noise Levels

Fig. 2: (a) Vision-based Offloading Strategy in Different Degree of Noise. (b) Kinematic Offloading Strategy Performance.

Fig. 3: Correlation Analysis between Joint Torque and Step-Wise Redundancy

Task Domain	Sequence Info		Action Ratio		Mean Attn Weight
Task Domain	$L$	$1/L$	$P_{red}$	$P_{crit}$	$W_{red}$	$W_{crit}$
Pick & Place	50	0.020	82.5%	17.5%	0.008	0.076
Drawer Opening	80	0.012	86.4%	13.6%	0.005	0.062
Peg Insertion	60	0.016	81.2%	18.8%	0.007	0.058

TABLE II: Quantitative Analysis of Attention Distribution and Action Redundancy

Method	Cloud-Side		Edge-Side		Total
Method	Lat.	Load	Lat.	Load	Lat.	Load
Edge-Only	–	–	782.5ms	14.2GB	$782.5\pm 28.5$ ms	14.2GB
Cloud-Only	113.8ms	14.2GB	–	–	$113.8\pm 15.6$ ms	14.2GB
SAFE(Vision-Based)	62.5ms	9.5GB	315.2 ms	4.7GB	$377.7\pm 26.2$ ms	14.2GB
RAPID (Ours)	83.5ms	11.8GB	139.4ms	2.4GB	$\bm{222.9\pm 11.4}$ ms	14.2GB

TABLE III: Performance Comparison of Edge-Cloud Collaborative Inference on Simulation Benchmarks.

Fig. 5: A Case of RAPID Framework Completing Real-World Tasks (Task Name: Pick up the Banana and Put it into the Blue Bowl)

Method	Cloud-Side		Edge-Side		Total
Method	Lat.	Load	Lat.	Load	Lat.	Load
Edge-Only	–	–	812.6ms	14.5GB	$812.6\pm 34.1$ ms	14.5GB
Cloud-Only	121.5ms	14.5GB	–	–	$121.5\pm 23.0$ ms	14.5GB
ISAR(Vision-Based)	68.3ms	10.2GB	345.8ms	4.3GB	$414.1\pm 21.8$ ms	14.5GB
RAPID (Ours)	91.2ms	12.1GB	148.5ms	2.4GB	$\bm{239.7\pm 15.7}$ ms	14.5GB

TABLE IV: Performance Comparison of Edge-Cloud Collaborative Inference on Real-World Environments.

Method	Cloud-Side		Edge-Side		Total
Method	Lat.	Load	Lat.	Load	Lat.	Load
w/o $\theta_{comp}$ (Acc.)	52.4ms	10.2GB	228.5ms	4.0GB	$280.9\pm 22.7$ ms	14.2GB
w/o $\theta_{red}$ (Torque)	45.2ms	8.5GB	270.4ms	5.7GB	$315.6\pm 32.4$ ms	14.2GB
RAPID (Ours)	83.5ms	11.8GB	139.4ms	2.4GB	$\bm{222.9\pm 11.4}$ ms	14.2GB

TABLE V: Ablation Study of Dual-Threshold Partitioning on LIBERO Benchmark.

Temporal Modeling for VLA Models

See all 19 papers

RoboECC: Multi-Factor-Aware Edge-Cloud Collaborative Deployment for VLA Models
Mar 21, 2026Zihao Zheng, Hangyu Cao, Jiayu Chen +6Edge-Cloud Collaborative Inference Efficient VLA Model Inference
School of Computer Science, Peking University, Beijing, China · School of Computer Science, South China University of Technology, Guangzhou, China · School of Artificial Intelligence, Beijing Normal University, Beijing, China +1
EcoVLA: Energy-Efficient Device-Edge Co-Inference for Vision-Language-Action Models under Real-Time Constraints
Aug 16, 2026Ao Zhou, Bo Dai, Le Yu +5Efficient VLA Model Inference Edge Inference
Beihang University, No.37 Xueyuan Road, Beijing, China · Beijing University of Technology, No.100 Pingleyuan Road, Beijing, China
How Fast Can I Run My VLA? Demystifying VLA Inference Performance with VLA-Perf
Feb 20, 2026Wenqi Jiang, Jason Clemons, Karu Sankaralingam +1Efficient VLA Model Inference On-Device Inference
NVIDIA Research

Mar 21, 2026J/K move · Enter open · S save

RoboECC: Multi-Factor-Aware Edge-Cloud Collaborative Deployment for VLA Models

Zihao Zheng, Hangyu Cao, Jiayu Chen, Sicheng Tian+5

School of Computer Science, Peking University, Beijing, China · School of Computer Science, South China University of Technology, Guangzhou, China · School of Artificial Intelligence, Beijing Normal University, Beijing, China+1

Edge-Cloud Collaborative Inference Efficient VLA Model Inference Edge-Cloud Computing Edge Computing Distributed Inference VLMs for Robotics

Open paper AskPDF arXiv

Vision-Language-Action (VLA) models are mainstream in embodied intelligence but face high inference costs. Edge-Cloud Collaborative (ECC) deployment offers an effective fix by easing edge-device computing pressure to meet real-time needs. However, existing ECC frameworks are suboptimal for VLA models due to two challenges: (1) Diverse model structures hinder optimal ECC segmentation point identification; (2) Even if the optimal split point is determined, changes in network bandwidth can cause performance drift. To address these issues, we propose a novel ECC deployment framework for various VLA models, termed RoboECC. Specifically, we propose a model-hardware co-aware segmentation strategy to help find the optimal segmentation point for various VLA models. Moreover, we propose a network-aware deployment adjustment approach to adapt to the network fluctuations for maintaining optimal performance. Experiments demonstrate that RoboECC achieves a speedup of up to 3.28x with only 2.55%~2.62% overhead.

Abstract

Temporal Modeling for VLA Models

Explore similar work

RoboECC: Multi-Factor-Aware Edge-Cloud Collaborative Deployment for VLA Models

EcoVLA: Energy-Efficient Device-Edge Co-Inference for Vision-Language-Action Models under Real-Time Constraints

How Fast Can I Run My VLA? Demystifying VLA Inference Performance with VLA-Perf