A new multi-robot control system called COMPASS lets a single natural-language command steer up to 1024 robots at once.
Researchers built COMPASS as a decentralized architecture where each robot runs its own spatial transformer that aggregates multi-hop messages from its neighbors into a compact learned "feedback token." That token, not raw sensor state, is what actually flows through the language channel between robots. In tests, this setup held cohesive flocking formations that tracked commanded intent, beating both a centralized frontier LLM policy and a version that relied on language-only communication between robots. It also generalized zero-shot to flocks 16 times larger than its training scale and to ambiguous instructions it had never seen before.
Most demonstrations of LLM-controlled robots fall apart once the group grows past a handful of units, because coordination overhead swamps whatever reasoning the model is doing. COMPASS's results suggest the fix isn't a bigger, smarter central brain but smarter local math: let each robot reason locally and pass along a compressed signal instead of raw sensor chatter. That's a different bet than the "bigger model, more context" approach most robotics-LLM work has leaned on.
It's still a research paper with simulated-scale numbers, not a warehouse full of drones; the real test is whether 1024 robots hold formation this well once battery life, latency, and physical collisions enter the picture.