Route by Kinematics, Act by Observation: Kinematics-Supervised Expert Routing in MoE-Augmented VLA

2026-07-29

AI Summary

This work addresses the inefficiency of expert routing in Mixture-of-Experts (MoE)-enhanced Vision-Language-Action (VLA) models during inference, which stems from the absence of kinematic signals present during training. To resolve this, the authors propose KinRT, a novel approach that leverages kinematic clustering of action trajectories to generate prototype labels for explicit supervision of the router during training. At inference time, KinRT enables efficient expert selection using only visual-language observations, effectively transferring kinematic knowledge from the action space to the observation space. This method establishes the first explicitly kinematics-supervised routing mechanism, creating an asymmetric yet highly effective bridge between training and inference. Experiments demonstrate that KinRT outperforms existing dense and MoE-based VLA models by 23.26% on the RoboTwin benchmark and achieves a 20.27% improvement on a low-cost, self-built DIYRobot platform, confirming its strong cross-platform generalization capability.

Abstract

While MoE augments VLA via expert specialization, router suffers from ineffective expert routing owing to the kinematic heterogeneity of actions across manipulation tasks and, even worse, the unavailability of the kinematic signals at inference time. In this work, we first observe that most semantically distinct manipulation tasks reduce to multiple kinematic archetypes. Motivated by this finding, we propose Kinematics-supervised explicit routing (KinRT), a new paradigm that shifts from implicit, observation-driven expert routing to explicit, kinematics-guided expert dispatching. Specifically, we perform kinematic clustering on action trajectories into multiple kinematically coherent groups, whose IDs serve as ground truth to supervise the training of the router; at inference time, the router dispatches experts only using visual-language observations, without any reliance on action kinematics. KinRT actually introduces an asymmetric bridging mechanism that distills the task kinematics from the action space in training into the observation space at inference. In addition, to assess KinRT's cross-platform generalization, we build an economical, Do-It-Yourself robot (DIYRobot) platform from scratch using 3D-print technology ($<$ 2,000USD). Extensive experiments demonstrate KinRT's superiority over both dense and MoE-featured VLAs by more than 23.26% on RoboTwin benchmark and 20.27% on our introduced DIYRobot platform. Our code and DIYRobot platform will be open-sourced.

Problem

• Mixture of Experts
• Visual-Language-Action Models
• Expert Routing
• Kinematic Heterogeneity
• Manipulation Tasks

Innovation

• Kinematics-supervised routing
• Mixture of Experts (MoE)
• Visual-Language-Action (VLA)
• Expert dispatching
• Cross-platform generalization