In swarm robotics, more robots cover more ground, but a map is only as complete as the areas robots have physically visited. SenseSwarm addresses this by having the swarm predict what lies ahead. The system runs a lightweight neural network on the swarm's shared occupancy map and feeds the predicted structure back into a stigmergic guidance field, so robots are drawn toward unvisited regions more efficiently.

The authors tested SenseSwarm on twelve real building floors up to 24,000 square meters, running 1,620 simulations across swarm sizes from 32 to 512. On physical coverage, it was roughly on par with a greedy frontier planner. But on map reconstruction, every non-predictive baseline needed 1.3 to 1.9 times longer to reach an 80% known map, and SenseSwarm led by about 670 square meters at equal time, up to 1,359.

A key design choice is that the prediction is computed once on the shared map rather than per robot. The authors found that applying the same predictor per robot matched SenseSwarm's speed but caused a 6.9 times higher inter-robot contact rate at 512 robots. Because the shared prediction is computed once, its computational cost stayed flat as the swarm scaled from 32 to 512 robots.