SOC in an aquatic robot swarm

The claim

A physical, programmable robot swarm engineered to exhibit self-organized criticality (SOC) — a gap prior platforms left open (ultracold-atom SOC is isolated with few tunable knobs; most robot swarms run on inert surfaces without dynamic physical coupling). Each aquatic robot is autonomous, and neighbours interact through the environment: optical attraction (phototaxis toward others’ LEDs — positive feedback) and hydrodynamic repulsion (water waves from vibrating films — negative feedback), forming a decentralised nonlinear feedback loop. The swarm shows the three hallmarks of SOC — power-law avalanches, finite-size-scaling universality, and spontaneous evolution to a steady state independent of parameters — and, under stimuli, retains criticality while producing adaptive collective behaviour.

Correlated clusters and power-law cascades

The avalanche observable is a correlated cluster: two individuals are directly correlated if within a distance $\varepsilon$ (= 11 cm) and with velocity directions within an angle $\theta$ (= 60°); the relation is transitive, so a cluster is a connected component and its size is $S=|C|$ . (This is essentially a proximity-plus-alignment contact graph.) For $N=64$ robots:

cluster size distribution $D(S)\sim S^{-\tau}$ , $\tau = 1.68(3)$ — close to neural avalanches (1.5), forest fires (1.4), atomic gases (1.37);
cluster duration $D(T)\sim T^{-\alpha}$ , $\alpha = 1.54(2)$ ;
size–duration relation $\langle T\rangle \sim S^{\gamma}$ , $\gamma = 0.74(4)$ ;
with a finite-size cutoff near the system size, and $1/f$ -type noise (spectral exponent $\beta\approx1.26$ ).

Finite-size scaling universality

A 2D lattice simulation (light modelled as an excitation field $\propto 1/r^2$ , water-wave as a negative field behind moving agents) run at $N = 100$ , $1000$ , $10{,}000$ (fixed density) keeps power-law cascades with essentially fixed exponents across a 100× size change. The simulation’s absolute exponents differ from the experiment (temporal discretisation, 8-direction motion, simplified collisions), but the authors stress that SOC universality means the ability to self-organise into a scale-invariant state — not exact exponent agreement, since exponents depend on microscopic details and observational definitions.

Self-organization to a steady state — with a coupling threshold

Unlike a tuned phase transition ( $T\to T_c$ ), the swarm evolves to criticality on its own; quadrupling light intensity and density still yields power-law cascades. Tracking the Shannon entropy $H = -\sum_S P(S)\log_2 P(S)$ of the cluster-size distribution reveals a threshold: below it (light $I<4$ cd or density $\rho<3.99\,\mathrm{m}^{-2}$ ) the excitation-field feedback network is too fragmented, entropy settles at scattered values, and the distribution is not power-law; above it, entropy converges to a fixed value independent of the parameters (≈3.57 bits for light, ≈3.44 for density) and the cascades are power-law. Criticality requires the coupling to exceed a threshold; beyond it the steady state is parameter-independent.

Robustness and emergent function

A passive infrared stimulus source (robots raise their own brightness with local IR strength) is not individually localisable, but aggregation produces a brightness gradient that biases a subcluster’s net motion toward it — yielding unprogrammed collective pushing (transporting the source at ~0.63 cm/s). Adding stimulus sources does not destroy criticality (cascades stay power-law); fixed sources act as spatial anchors, and two of them induce a spontaneous “bridge” structure between clusters. Function emerges from local interactions and the underlying SOC state, without global control — the authors frame this as a route to adaptive swarm robotics with minimal onboard computation.

References

Zhao et al. (2026), Sci. Adv. 12, eaec6153 · Bak, Tang & Wiesenfeld (1987), Phys. Rev. Lett. 59, 381 · Cavagna et al. (2010), PNAS 107, 11865 · Shew & Plenz (2013), The Neuroscientist 19, 88 · Rubenstein, Cornejo & Nagpal (2014), Science 345, 795 · Marković & Gros (2014), Phys. Rep. 536, 41 · Klaus, Yu & Plenz (2011), PLoS ONE 6, e19779 · Marshall et al. (2016), Front. Physiol. 7, 250.