Latency-Tolerant Cloud-Edge Collaborative Vision-Language-Action Models via Emergent Representational Specialization

2026-08-01

AI Summary

This work addresses the system conflict arising from latency and jitter between cloud-based semantic reasoning and local real-time control when deploying billion-parameter vision-language-action (VLA) policies on mobile robots. The authors propose CloudEdgeVLA, the first framework to model temporal misalignment in asynchronous cloud-edge observations as a representation learning problem: the cloud encodes slowly varying task semantics, while the edge fuses the latest cloud features with current visual inputs to enable low-latency control without explicit scheduling. The approach employs a cloud-edge协同 architecture augmented with random-delay frame training, dual-path (fresh/stale) action prediction, and a lightweight edge head. Evaluated on four LIBERO benchmarks under uniform 40-step delays, CloudEdgeVLA achieves success rates of 63.8%–78.0%, substantially outperforming VLASH (≤6.4%) and single-path baselines (≤3.0%).

Abstract

Deploying billion-parameter Vision-Language-Action (VLA) policies on mobile robots creates a systems conflict: semantic reasoning benefits from cloud GPUs, whereas closed-loop control must respond locally despite network delay and jitter. Existing hierarchical and asynchronous policies improve throughput, but their slow-path representations can still arrive stale or require explicit scheduling and delay cues. We introduce CloudEdgeVLA, a cloud-edge policy that treats temporal misalignment as a representation-learning problem. A cloud VLA encodes delayed observations into slowly varying task features, while a lightweight edge head combines the latest available cloud feature with current local vision. During training, current and randomly delayed frames are paired with the same current action target in fresh and stale paths. This objective encourages the cloud representation to preserve task-level information while the edge path supplies state-sensitive corrections. Across four LIBERO suites, CloudEdgeVLA retains 63.8--78.0% success with a 40-step uniform-delay window, whereas VLASH reaches at most 6.4% and the evaluated single-path baselines at most 3.0%. By removing blocking synchronization from the control loop, the design offers a practical route to scalable VLA deployment in which cloud models can grow while edge computation remains lightweight and responsive.

Problem

• Vision-Language-Action
• Cloud-Edge Collaboration
• Latency Tolerance
• Temporal Misalignment
• Mobile Robotics

Innovation

• cloud-edge collaboration
• vision-language-action models
• latency tolerance
• representational specialization
• asynchronous control