COMPASS Lets Language Models Steer Swarms of 1,024 Robots

A team of researchers has proposed COMPASS, a decentralized architecture designed to solve a persistent problem with using large language models (LLMs) for multi-robot control: coordination tends to break down as the number of robots increases. The work is described in a paper posted to arXiv (arXiv:2609.28247).

Instead of relying on a single centralized model to direct an entire fleet, COMPASS runs feedback generation locally on each robot. A spatial transformer on each unit aggregates multi-hop messages passed across the fleet and compresses them into a learned feedback token, which is then used to steer each robot's reasoning process.

The researchers found that giving individual agents structured diversity in their input commands actually improved performance, helping cancel out biases — a benefit that held even as the number of robots scaled up. In head-to-head comparisons against a centralized frontier LLM policy and a language-only communication baseline, COMPASS produced markedly more cohesive flocking formations that better matched the intent behind natural-language commands.

Ablation studies revealed an important design detail: reasoning feedback worked best when paired with a compact learned token rather than raw data. When the team instead fed hand-engineered feedback containing raw robot state directly into the language channel, flock cohesion collapsed entirely.

Perhaps the most striking result is COMPASS's scalability. The system generalized zero-shot to new instructions with ambiguous phrasing and successfully commanded flocks up to 16 times larger than the scale it was trained on, flying formations of up to 1,024 robots under natural-language instructions.

As reported in the paper, this points toward a path for using LLMs not just for planning single robots but for directing genuinely large collectives — a capability that has so far eluded existing approaches. Since the work is currently a preprint, it has not yet undergone formal peer review, and details such as the physical robot platform used in testing were not specified in the abstract.