A new training method teaches robot teams to coordinate instead of just working side by side.
Researchers built a three-stage reinforced fine-tuning pipeline for vision-language-action models, the systems that let robots translate a camera feed and a text instruction into physical actions. The first stage collects new training data only when a pretrained model keeps failing, cutting down on costly human demonstrations. The second stage tunes each robot's model separately, crediting only the trajectory segments that actually helped, rather than treating a whole multi-robot rollout as one unit. The third stage runs reinforcement learning in the model's latent space while keeping the base model frozen, a workaround for what the team found was unstable, noisy learning when robots explore jointly online.
That matters because most vision-language-action models are pretrained on single-robot data and have no built-in sense of how to hand off a task or avoid working at cross purposes. Tested on the pi0 and pi0.5 model backbones across 11 tasks in the RoboTwin and RoboFactory simulators plus real manipulation with two Franka robots, the method lifted average success rates by 23.1 percent, 16.4 percent, and 44 percent respectively. For warehouses and factories eyeing multi-robot automation, that is the difference between machines that merely coexist and ones that actually divide labor.
The real-world test used exactly two robots and code posted to an anonymous repository, so treat the headline numbers as promising lab results, not a fleet-ready product.