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

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.

💡 Research Summary

The paper tackles the high inference cost of Vision‑Language‑Action (VLA) models, which are increasingly used for embodied intelligence tasks such as robot manipulation. While edge‑cloud collaborative (ECC) inference is a promising system‑level solution, existing ECC frameworks rely on environment‑oriented, vision‑based partitioning that suffers from two fundamental drawbacks. First, they use visual confidence metrics (e.g., Shannon entropy of the model’s action distribution) to decide when to offload computation to the cloud. This makes the decision process highly sensitive to visual noise, lighting changes, background clutter, and camera motion, causing unnecessary cloud offloads and large latency spikes. Second, VLA models exhibit step‑wise redundancy: many consecutive time steps correspond to “unimportant” actions (e.g., smooth approach motions) that contribute little to the overall task, while a few steps involve critical interactions that demand high‑capacity processing. Existing partitioning schemes ignore this redundancy, leading to inefficient use of edge and cloud resources.

RAPID (Redundancy‑Aware and Compatibility‑Optimal Edge‑Cloud Partitioned Inference for Diverse VLA Models) is proposed to overcome these issues. Its design rests on two complementary mechanisms:

1. Compatibility‑Optimal Partitioning – Instead of visual cues, RAPID monitors proprioceptive kinematic signals directly available on the robot: joint acceleration (¨q) and joint torque (τ). The instantaneous joint acceleration magnitude M_acc(t) = ‖W_a·¨q_t‖₂ is computed at each control step, where W_a is a diagonal weight matrix emphasizing joints that are more informative for task changes. A sliding window maintains the mean µ_acc and standard deviation σ_acc of M_acc; the normalized anomaly score ˆM_acc(t) = (M_acc(t) – µ_acc) / (σ_acc + ε) is compared against a threshold. Abrupt, non‑linear motions such as sudden stops, direction changes, or collision avoidance generate spikes in this score, triggering a cloud offload. Because the metric derives from the robot’s own dynamics, it is invariant to external visual disturbances, ensuring consistent partitioning across diverse environments.

2. Redundancy‑Aware Partitioning – An analysis of internal attention weights in VLA models shows that critical interaction steps have high attention values, while the majority of steps (≈80 % in the evaluated tasks) have very low attention, indicating redundancy. Directly extracting attention is computationally expensive, so RAPID uses joint torque as a lightweight surrogate. Empirical correlation studies demonstrate that torque spikes align tightly with attention peaks during critical interactions. Consequently, RAPID classifies a time step as “high‑redundancy” when both torque magnitude and M_acc are low, keeping the computation on the edge, and as “low‑redundancy” when either metric spikes, offloading the step to the cloud. This dual‑threshold approach balances the need for real‑time edge control with the cloud’s capacity for complex reasoning.

The authors implement RAPID on a 6‑DOF manipulator equipped with low‑latency joint sensors and a remote cloud server. They evaluate three representative manipulation tasks—Pick‑and‑Place, Drawer Opening, and Peg Insertion—under four visual conditions: clean, varying illumination, background noise, and visual distraction. Baselines include Edge‑Only VLA, a vision‑based entropy trigger (ISAR), and static layer partitioning. Results show that RAPID achieves up to 1.73× speedup while incurring only 5 %–7 % additional communication overhead. In noisy visual settings, the vision‑based method’s offload frequency explodes, leading to up to 2× higher total latency, whereas RAPID’s offload rate remains stable. Moreover, task success rates improve by up to 15.8 %, reflecting more accurate handling of critical interaction phases.

The paper’s contributions are threefold: (1) Demonstrating that kinematic features are robust against visual noise and correlate strongly with step‑wise redundancy; (2) Proposing a dual‑threshold ECC framework that leverages these features for both compatibility and efficiency; (3) Providing extensive empirical evidence across multiple tasks and noise levels. Limitations include dependence on accurate joint torque sensing (which may not be available on all platforms) and evaluation primarily in simulated or controlled lab environments. Future work is suggested on extending the approach to heterogeneous robot morphologies, integrating adaptive threshold learning (e.g., reinforcement learning) to handle dynamic network conditions, and exploring multi‑robot collaborative scenarios where shared kinematic cues could further optimize cloud utilization.

In summary, RAPID introduces a novel, motion‑centric ECC paradigm that replaces fragile visual triggers with physics‑based signals and exploits intrinsic redundancy in VLA inference. This yields a system that is both environment‑agnostic and computationally efficient, marking a significant step forward for deploying large‑scale multimodal models in real‑time embodied AI applications.