Vision-language-action (VLA) models are expanding what individual robots can do, but many real-world manipulation settings require multiple arms to operate as a coordinated system. Existing approaches organize multi-arm coordination through centralized joint policies, high-level task or role assignment, or local policies that react to teammates. The authors of VLA-Fabric note that how complete action-producing VLAs should explicitly interact with one another during action generation has remained largely unexplored.
VLA-Fabric gives each arm its own complete VLA, preserving local perception and action ownership, and enables coordination through learned interaction across policy boundaries. The researchers systematically studied where interaction should occur, how peer information should be aggregated, and how the interaction should be learned, revealing complementary coordination functions: shared context, peer-specific latent information, and action-stage interaction. They then re-instantiated these functions in a structurally different VLA architecture, showing that the coordination principle is not tied to a particular tensor layout or backbone implementation.
In experiments with teams of three and four arms, in both simulation and physical tests, full interaction improved success over independently trained local policies by 16 to 40 percentage points on all four physical tasks. The authors suggest that explicit interaction among complete VLA policies offers a scalable path toward networked embodied intelligence.