MoE-VLA: MoE Pathways for Efficient Vision-Language-Action

Abstract

Efficient deployment of vision-language-action (VLA) policies is difficult because closed-loop robot control repeatedly invokes multimodal inference at every control step. This paper studies Mixture-of-Experts Vision-Language-Action (MoE-VLA), a lightweight Mixture-of-Experts (MoE) routing framework for efficient VLA inference. The key idea is to treat layer-execution paths as computation experts and select a path from action change, visual variation, and temporal context. Instead of using a fixed pruning rule, MoE-VLA adaptively combines cache reuse, visual-token reduction, and layer skipping. To keep the mechanism easy to analyze, we construct a controlled Synthetic Mini-VLA benchmark with four tabletop manipulation task groups and evaluate representative efficiency strategies on the same compact backbone. MoE-VLA further uses an on-policy-distillation-style sparse-path objective to preserve dense-policy behavior under routed execution. Experiments show that MoE-VLA achieves a favorable successefficiency trade-off across Visual Matching and Variant Aggregation settings, reaching 83.1% and 61.0% average success with 2.55× and 2.29× effective speed-up, respectively. These results indicate that MoE-style path routing can reduce redundant computation while preserving task-relevant control behavior.