Robotics study: Better communication beats bigger AI models in multi-robot tasks
Restructuring robot chat nets 47-point gain vs 9 for doubling model size
A new arXiv paper from Junping Wang, Zhizhong Zhang, and colleagues investigates a fundamental question in multi-robot coordination: given a fixed hardware budget, is it better to increase each robot's onboard AI model size or restructure how they communicate? The team ran a real-world transport-and-mapping task using 10 physical robots across 60 trials (5 runs per condition). They compared a fully connected communication topology against a modular hierarchical structure, and separately tested doubling the hidden size of each robot's neural network.
The results are striking. Switching to the hierarchical communication structure improved normalized performance by 47 points on a 0–100 scale, whereas doubling hidden layer size yielded at most 9 points improvement. Nested mixed-effects models confirmed a substantially larger fit improvement from topology changes than from model scaling. The pattern held in independent replications using the SMAC benchmark, and simulation-calibrated extrapolation showed performance saturation beyond 1,024 hidden units. The authors note broader quantitative generalization remains to be established, but within this realistic task, communication structure dominated scaling.
- Modular hierarchical communication improved performance by 47 points vs only 9 points from doubling neural network hidden size
- Results come from a real-world experiment with 10 physical robots and 60 total runs across conditions
- Simulation-based extrapolation showed performance saturation beyond 1,024 hidden units, suggesting diminishing returns from larger models
Why It Matters
For teams deploying robot swarms: smarter coordination structures may be more cost-effective than bigger AI chips.