Event

SenseSwarm proposes a swarm mapping method based on shared expectations

1 reports · 1 sources

Haohua Que et al. published a paper on arXiv proposing the SenseSwarm method: a lightweight 709K-parameter map predictor runs on the shared occupancy map of a robot swarm, folds the predicted structure back into a stigmergy-based guidance field, and fine-tunes the predictor using the swarm's own multi-frontier observations—the paper notes that a predictor trained on a single robot fails under this perspective. The paper released a benchmark composed of real building floors from HouseExpo, HKUST, and KTH (12 floors in total, up to 24,000 square meters), and ran 1,620 simulations covering 9 collaborative methods, 32 to 512 robots, and 3 random seeds. Results show that the swarm is roughly on par with a greedy frontier planner in physical coverage (the paper states this is as expected); the prediction is computed only once per shared map, and its overhead remains constant from 32 to 512 robots.