Vision language action models coordinate multiple robot arms effectively
VLA-Fabric: Learning Interaction for Composable Multi-Arm VLA Agents
Robotics
Summary
Controlling multiple robot arms to work together is hard because each arm needs to understand what the others are doing. The authors studied how these robot arms can share information and cooperate better while still making their own decisions. They developed a system called VLA-Fabric that lets each arm use its own vision-language-action model but interact with other arms during task execution. Their experiments showed that this teamwork approach helps robots complete physical tasks more successfully than if each arm worked independently.
What this means in practice
- •For robotics engineers: Coordinate multiple robot arms in manufacturing or assembly lines with improved success from learned interaction policies.
- •For automation system developers: Develop multi-arm robotic systems with scalable frameworks for synchronized manipulation tasks in warehouses or labs.
Authors
Chenglong Zhang, Shuaijun Liu, Feiyang You, Shuyang Hao, Chengyu Wu, Ningxin Su
Abstract
Vision-language-action (VLA) models are rapidly expanding what individual robots can accomplish, yet many real-world manipulation settings require multiple arms to operate as a coordinated system. Existing approaches organize multi-arm coordination in different ways, including centralized joint policies, high-level task or role assignment, and locally executed policies that react to teammates. However, how complete action-producing VLAs should explicitly interact with one another during action generation remains largely unexplored. We introduce VLA-Fabric, which organizes one complete VLA per arm while preserving local perception and action ownership, and enables coordination through learned interaction across policy boundaries. We first systematically study where such interaction should occur, how peer information should be aggregated, and how the interaction should be learned, revealing complementary coordination functions spanning shared context, peer-specific latent information, and action-stage interaction. We then re-instantiate these functions in a structurally different VLA architecture, showing that the coordination principle is not tied to a particular tensor layout or backbone implementation. Finally, we compose teams of three and four arms and evaluate them in simulation and physical experiments. On all four physical tasks, Full interaction improves success over independently trained local policies by 16--40 percentage points. Together, these results suggest that explicit interaction among complete VLA policies offers a scalable path toward networked embodied intelligence.