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 target the amount of an unknown building a decentralized swarm knows at each instant, ahead of physical coverage. 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, and SenseSwarm holds an equal-time area lead of about 670 square meters, up to 1,359. 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.
Figures & tables
Fig. 1: The SenseSwarm loop (top) and the shared map service (bottom). Robots sense locally into the shared occupancy map K . The service tiles the frontier of K with 50 percent overlap, runs the swarm-distribution predictor once per shared map, and publishes two products of the predicted structure σ : the predicted free cells P(t) folded with the sensed map S(t) into the reconstruction R(t) , and the uncertainty u=1−2∣σ−0.5∣ that seeds the predictive field Φ at the frontier. Each robot reads only the four field gradients at its own position and steers by Eq. ( 7 ) under reactive avoidance, so the per-robot cost is independent of the swarm size.
1:
Shared map service:
2:
integrate all newly sensed cells into occupancy grid K
3:
update frontier-distance field F (BFS from frontiers)
4:
decay and deposit explore trace E , density trace ρ
5:
if tmodTpred=0 then
6:
tile the frontier neighborhood of K ; batch-predict σ
7:
u←1−2∣σ−0.5∣ on predicted unknown cells
Table 2
Fig. 2: The twelve benchmark floors, individually scaled, walls ink and free space white: three HouseExpo apartments (a to c, 2,439 to 9,574 m 2 free), four KTH corridor composites of real stories (d to g, 3,752 to 14,870 m 2 ), and five HKUST campus floors from the SLABIM building information model (h to l, 10,376 to 24,015 m 2 ). Per-floor areas are in Table I ; full-resolution maps are on the project page.
Fig. 3: Coverage versus time on all twelve floors, median over three seeds at 256 robots: thin lines are the six completing baselines’ sensed coverage (the settling gradient caps and is omitted), the thick purple line is SenseSwarm’s reconstruction, and the shaded band above the best baseline is the building known but not yet visited, annotated as the time-averaged lead in points and square meters.
Fig. 4: Time to an 80 percent known map per floor, normalized so SenseSwarm is 1.0, at 256 robots. Bars are the mean over three seeds and error bars the seed standard deviation; the Mean group averages the twelve floors. Hatched bars marked cap never reach 80 percent. Every non-predictive baseline is above the purple dashed line on nearly every floor: nearest-frontier and MMPF need about 1.3 times our time, the DMPF-style field 1.5, and MinPos 1.9. Pure stigmergy is SenseSwarm without prediction, so its 1.31 bar is the ablated cost of removing the predictor. The per-robot integration ablation is excluded here; it ties our speed and Fig. 5 c shows its safety cost.
Fig. 5: Scaling and safety, medians over the twelve floors. (a) Normalized time to an 80 percent known map versus swarm size: every baseline holds a flat margin above SenseSwarm from 32 to 512 robots. (b) Per-floor reconstruction gain versus floor area with a 1/area guide (Spearman r=−0.62 ): the gain shrinks dimensionally with area, and structure modulates the trend (HK-4 versus KTH-4). (c) Contact rate versus swarm size, log-log with interquartile bands, clipped one decade below the median where the lower quartile is zero: per-robot prediction (dashed) congests like uncoordinated nearest-frontier; SenseSwarm stays an order of magnitude lower. (d) Speed versus safety at the harder 90 percent known threshold (a robustness check beyond the 80 percent used elsewhere): time to a 90 percent known map against the contact rate at 512 robots; SenseSwarm alone occupies the fast-and-safe corner.
Floor
Free (m 2 )
Faster
Lead (pp)
Lead (m 2 )
kth_single
3,752
+37%
+13
+487
he_small
2,439
+33%
+13
+320
kth_double
5,359
+29%
+13
+685
he_medium
5,739
+24%
+10
+557
hkust_4f
23,028
+19%
+6
+1,359
hkust_3f
21,869
+14%
+5
+1,055
TABLE I: Gain per floor versus the per-run best baseline: time-to-80%-known reduction and equal-time lead.
Floor
NF
MMPF
MinPos
DMPF
Stig
Lane
Set
Ours
HE-S
1.52
1.64
1.96
1.99
1.63
1.64
cap
1.00
HE-M
1.37
1.47
2.26
1.72
1.46
1.49
cap
1.00
HE-L
1.13
1.23
2.01
1.39
1.20
1.21
cap
1.00
HK-1
1.14
1.17
cap
1.37
1.22
1.20
cap
1.00
HK-2
1.32
1.07
1.62
1.19
1.22
1.09
cap
1.00
HK-3
1.19
1.21
cap
1.36
1.21
1.21
cap
1.00
TABLE II: Time to an 80% known map per floor at 256 robots (ours = 1.0; cap = never). NF = nearest-frontier, Stig = pure stigmergy, Lane = handed-lane, Set = settling. Bottom: Wilcoxon against each baseline.
Method
32
64
128
256
512
Time to 80% known (norm.)
nearest-frontier
1.29
1.25
1.27
1.28
1.29
MMPF
1.28
1.32
1.31
1.31
1.33
MinPos
2.26
2.08
2.01
1.93
1.94
DMPF-style PF
1.50
1.48
1.47
1.49
1.55
pure stigmergy
1.40
1.34
1.31
1.31
1.31
TABLE III: Method by swarm size: time to 80% known and 95% sensed (ours = 1.0; cap = never), contacts per robot per 100 m, travel (m/m 2 ), and blocked steps (%). The per-robot ablation ties our speed by construction and pays in contacts.
Floor
Metric
full
no discount
area seeding
he_medium
time (256)
1.00
0.96
1.00
he_medium
time (512)
1.00
1.00
1.01
he_medium
contacts
0.50
0.59
0.63
kth_single
time (256)
1.00
0.97
0.97
kth_single
time (512)
1.00
1.06
1.07
kth_single
contacts
1.93
1.25
1.62
TABLE IV: Component ablations on he_medium and kth_single: guidance design (top) and sensing range (bottom).
Fig. 6: Coordination methods at 256 robots: the same apartment floor (top) and campus scale, the 23,000 square meter hkust_4f (bottom). Path colors match the method colors of Figs. 3 to 5 and darken with time; SenseSwarm is purple. Chips state the known map fraction, which for the non-predictive baselines equals their sensed coverage; predicted structure is blue, and unknown space carries a faint ground-truth outline; nothing faint is visible to any method. Trails link logged positions and omit chords that would cross a wall between samples. Baselines bunch along corridors; SenseSwarm disperses toward predicted rooms and leads nearest-frontier’s known map by 9 to 15 points at equal campus ticks.
Fig. 7: Sensed map versus reconstruction over time: a HouseExpo apartment floor at 256 robots (top block) and the twelve-story kth_quad composite, 231 meters long, at 512 robots (bottom block), the hardest floor for prediction, whose stories connect through a few shafts crossed serially (Sec. VII ). SenseSwarm paths are purple and darken with time, predicted structure is blue, and unknown space carries the faint ground-truth outline. Rooms appear in the reconstruction several panels before any robot enters them, and even the shaft-serialized composite stays 4 to 8 points ahead.
Observation distribution
obstacle IoU
single-robot original
0.34
swarm fine-tune, obstacle-weighted BCE
0.73
swarm fine-tune, + soft clDice (ours)
0.85
Generalization (held-out hkust_3f)
gain
paired predictor, floor in training
+14.4%
leave-one-building-out, floor never seen
+10.8%
TABLE V: Predictor ablations: obstacle IoU on the held-out swarm split (top); leave-one-building-out gain on hkust_3f (bottom).
Department of Electronic Engineering, Tsinghua University, Beijing, China. · Tsinghua Shenzhen International Graduate School (SIGS), Tsinghua University, Shenzhen, China. · School of Software and Microelectronics, Peking University, Beijing, China.