SenseSwarm: Swarm Exploration via Neural Structure Estimation with Shared World-Anticipation for Rapid Mapping

Anonymous Authors

Under double-blind review

📄 Paper (PDF) 🎥 All 108 Runs 🏠 Benchmark 📚 BibTeX

512 robots explore a real 231-meter, twelve-story composite building. The purple swarm reconstructs the map (blue = predicted structure) ahead of physical coverage: the chip counts what the swarm knows, not only what it has visited.

Abstract

A robot swarm covers a building faster with more robots, but the map is only as complete as the space the robots have physically visited. We ask a different question: how much of an unknown building can a decentralized swarm know, at each instant, before it has walked through it? SenseSwarm serves a lightweight learned map predictor on the swarm's shared occupancy map and folds the predicted structure back into a stigmergic guidance field. Its 709K-parameter predictor is fine-tuned on the swarm's own multi-front observations, which we show is necessary: a single-robot predictor collapses on such views. We build a benchmark of twelve real building floors from HouseExpo, HKUST, and KTH, up to 24,000 square meters, and run 1,620 simulations across nine coordination methods, swarm sizes 32 to 512, and three seeds. On physical coverage the swarm is near parity with a greedy frontier planner, as expected. On map reconstruction, every non-predictive baseline needs 1.3 to 1.9 times longer to reach an 80 percent known map. Folding the prediction into the shared field matters: the same predictor applied per robot matches our speed but suffers a 6.9 times higher inter-robot contact rate at 512 robots. The prediction is computed once per shared map, so its cost stayed flat from 32 to 512 robots.

Key idea: predictive stigmergy. One predictor runs on the shared map, and its output becomes an uncertainty-seeded, density-discounted stigmergic field that every robot reads only locally. The cost tracks the frontier, not the robot count. The value is not walking the building faster; it is knowing the building before you walk it.

Method

The SenseSwarm loop. Robots sense locally into the shared occupancy map K. A shared map service tiles the frontier (50% overlap), runs the swarm-distribution predictor once per map, and publishes the reconstruction R(t) and the shared fields F, E, ρ, Φ. The prediction enters coordination as a fourth stigmergic field: uncertainty u = 1 − 2|σ − 0.5| seeds the frontier, and each robot steers by four local gradient reads under reactive avoidance, so the per-robot cost is independent of the swarm size.

Reconstruction leads physical coverage

Coverage versus time on all twelve floors (median over three seeds, 256 robots). The purple curve is SenseSwarm's reconstruction; the shaded band above the best baseline is the building known but not yet visited.

Time to an 80 percent known map, normalized so SenseSwarm is 1.0. Every non-predictive baseline is 1.3 to 1.9 times slower on average; the ordering holds on nearly every floor and is significant at p ≤ 0.002 against each baseline (Wilcoxon signed-rank).

(a) The margin is flat from 32 to 512 robots. (b) The gain shrinks dimensionally with floor area. (c) Contact rates diverge as the swarm densifies; per-robot prediction congests like uncoordinated frontier. (d) SenseSwarm alone occupies the fast-and-safe corner.

Qualitative behavior

The same floor, the same clock, side by side. Physical coverage stays close: a greedy frontier planner is already near optimal at walking a building. What pulls ahead is the known map, because SenseSwarm reconstructs rooms (blue) before any robot enters them.

Apartment floor (HE-M), 256 robots

Campus floor (HK-4, 23,000 m²), 256 robots

Twelve-story composite (KTH-4, 231 m), 512 robots

The benchmark: twelve real floors

Three HouseExpo apartments, four KTH corridor composites built from real stories by a deterministic seeded rule, and five HKUST campus floors rasterized per BIM category from SLABIM, up to 24,015 square meters. Full-resolution maps are in benchmark_maps/ and the construction scripts in scripts/.

Every floor, every method

All 108 exploration runs (12 floors × 9 methods, 256 robots, seed 1), replayed from the logged campaign. Pick a floor:

Each clip plays its own run, so at any moment the panels are at different ticks; for equal-time comparison use the shared-clock videos above. Per-robot prediction is our own ablation, not a competing method: it runs the same predictor on every robot, which is why it matches our speed. It pays for that speed in collisions, 3.58 against our 0.52 contacts per robot per 100 m at 512 robots.

BibTeX

@inproceedings{senseswarm,
  title     = {SenseSwarm: Swarm Exploration via Neural Structure Estimation
               with Shared World-Anticipation for Rapid Mapping},
  author    = {Anonymous},
  booktitle = {Under review},
  year      = {2026}
}

Anonymous project page for double-blind review. Website layout adapted from the Nerfies template (CC BY-SA 4.0).