Robot coordination: communication structure beats model scaling
Research published on arXiv (2605.30383) indicates that enhancing communication methods among robots leads to better coordination than merely increasing the size of onboard models. In an experiment involving 10 physical robots (with 5 trials for each condition, totaling 60 trials), transitioning from fully connected to modular hierarchical communication resulted in a 47-point increase in normalized performance (on a 0–100 scale). In contrast, enlarging the neural network's hidden size produced a maximum improvement of only 9 points. Comparisons using nested mixed-effects models demonstrated a significantly greater enhancement in model fit due to topology rather than scale. This trend was also observed in separate SMAC simulations, and heterogeneous benchmark reanalyses provided additional validation. Performance plateaued after exceeding 1024 hidden units.
Key facts
- Study on arXiv 2605.30383
- 10 physical robots used
- 5 runs per condition, 60 runs total
- Modular hierarchical interactions improved performance by 47 points
- Doubling hidden size yielded at most 9 points
- Nested mixed-effects model comparisons showed larger improvement for topology
- Pattern confirmed in independent SMAC replications
- Performance saturation beyond 1024 hidden units
Entities
Institutions
- arXiv