Robot swarms look very different when they jump from simulation to real‑world stages
arXiv: 2611.05244v1
In a series of live art performances, researchers moved the identical swarm‑control code to three distinct robot teams—tiny tabletop “Sushis”, flying CrazyCognies drones, and human‑scale mobile manipulators called Doodies. The software stayed identical, but the robots moved at vastly different speeds and traced different shapes. On the aerial drones, one cycle of an executive choreography took almost 27 seconds, while the same command ran in less than 8 seconds on the ground robots and 12 seconds on the tabletop machines. The drones’ motion patterns were also less sharply separated from other behaviours, showing that the same code can yield very different visible behaviour when the underlying physics change.
Table 1 summarizing per‑platform measurements of radius of gyration, radial index, polarization, milling order, and modulation depth for each behavior, illustrating differences in observable characteristics across the three platforms.
Doody
CrazyCogni
Sushi
Fireworks
Median Rg (m)
1.45±0.24
1.53±0.05
0.166±0.004
Median speed (mm s-1)
130±40
166±25
34.8±1.0
ψ
+0.57±0.07
+0.16±0.07
+0.653±0.015
P
0.33±0.08
0.29±0.07
0.157±0.015
M
0.175±0.029
0.17±0.04
0.055±0.008
Cyclic pursuit
Median Rg (m)
1.4±0.4
1.601±0.008
0.1061±0.0029
Median speed (mm s-1)
151.4±2.0
153.2±3.3
85.3±3.2
ψ
−0.61±0.22
−0.19±0.05
−0.896±0.012
P
0.4±0.4
0.293±0.021
0.11±0.06
M
0.7±0.4
0.45±0.06
0.9768±0.0034
Lennard–Jones
Median Rg (m)
1.469±0.015
1.48±0.13
0.04137±0.00034
Median speed (mm s-1)
40±4
148±7
0.30±0.04
ψ
−0.023±0.035
−0.06±0.05
+0.05±0.07
P
0.376±0.017
0.47±0.08
–
M
0.190±0.021
0.158±0.026
–
Behaviour-specific
Fireworks cycle period (s)
7.58±0.09
26.9
12.036±0.023
Fireworks modulation depth
0.27±0.06
0.79
1.335±0.027
Pursuit orbital period (s)
58±18
101±10
7.87±0.07
Table 1. Rg, radius of gyration; ψ, radial index; P, polarization; M, milling order; modulation depth, peak‑to‑trough swing of the phase‑folded Rg over its median. Measurements from motion capture and overhead video show that identical behaviors produce markedly different observable characteristics across the three platforms—execution timescales vary by more than a factor of three, directional separation (ψ) is compressed on the CrazyCognies, and trajectory scales range from centimetre‑scale to metre‑scale orbits.Imran, Ali, Abdalwhab, Abdalwhab Bakheet Mohammed, Braga, Rafael Gomes, Duval, Hélène, Rochette, Audrey, Varadhajan, Vivek Shankar et al. · arXiv:2611.05244v1 · Table 1: · CC BY 4.0
Even though the on‑board computers calculated the same relative positions for the Swarm, the timing of each control step depended on platform hardware. The drones performed each cycle with roughly a 3.5‑fold lower controller frequency than the ground robots, which explains the 3.5‑fold spread in observed cycle times.
Observers who watched the performances did not agree that the visible differences matched the underlying technical success. A survey after each show found that people were more likely to rate the tabletop Sushis as friendly and meaningful, and less likely to perceive threat from them, than they were for the drones or the Doodies. In contrast, the drones were often seen as more threatening, even when the underlying algorithm behaved the same.
Localization accuracy also varied by platform. The drones’ horizontal pose estimates, derived from a UWB network, had a median error of about 0.08 m and only 1–3 % of samples were further than 0.5 m off. On the ground robots, the median error rose to roughly 0.11–0.15 m, with 2–6 % of samples exceeding 0.5 m. The larger errors on the ground vehicles could contribute to the lower perceptual appeal.
Table 2 reporting median horizontal localisation error (with interquartile range) and the percentage of samples exceeding 0.5 m for CrazyCognies and Doodies across the three behaviors.
Platform
Behaviour
n
Horizontal error (m)
Samples >0.5 m
median [Q1, Q3]
(%)
CrazyCognies
Fireworks
13
0.077 [0.049, 0.115]
1.1
Cyclic pursuit
15
0.077 [0.050, 0.111]
0.2
Lennard–Jones
13
0.078 [0.048, 0.111]
3.0
Doodies
Fireworks
14
0.128 [0.081, 0.224]
6.0
Cyclic pursuit
14
0.107 [0.065, 0.155]
2.1
Lennard–Jones
15
0.154 [0.098, 0.224]
3.1
Table 2. Localisation uncertainty during expressive behaviors. Horizontal position error of the onboard estimate relative to motion capture, by platform and behavior. Nine takes were recorded per platform, three per behavior. n gives the number of robot‑takes passing the correspondence criterion. Position error is reported as median [Q1, Q3]. The final column shows the fraction of synchronized samples with error exceeding 0.5 m.Imran, Ali, Abdalwhab, Abdalwhab Bakheet Mohammed, Braga, Rafael Gomes, Duval, Hélène, Rochette, Audrey, Varadhajan, Vivek Shankar et al. · arXiv:2611.05244v1 · Table 2: · CC BY 4.0
These results show that simply porting swarm software to new hardware does not guarantee that the swarm will behave the same to humans. Adjustments are needed not only to compensate for the “reality gap” of hardware and communication differences but also to address the perceptual gap that shapes how people read collective motion.
[Visual 1: Table 1 showing execution times, radial index, and trajectory scales across the three platforms]
[Visual 2: Table 2 showing localisation uncertainty for the drones and ground robots]