When robots in a swarm crash, the survivors have to decide which jobs to drop. Usually that decision is made by a central planner, by explicit messages, or by a learned policy that watches its teammates. This note proposes a fourth option that needs none of these. The shared environment holds one field per task, and the fields compete with each other the way superconductivity and charge-density waves compete for the same electrons in a metal. A charge-density wave is a periodic ripple of electron density; superconductivity is the pairing of electrons into a zero-resistance state. Because both draw on one pool of electrons, the stronger order takes them first and pushes the weaker one out.
In the swarm version, agents are the shared pool. Each agent writes its commitment into the fields where it stands and reads the fields to choose its next job. The fields themselves follow textbook Ginzburg–Landau competition rules. A Ginzburg–Landau model describes how competing "orders" grow and suppress each other according to a simple energy function. The payoff is that the swarm's response to losing agents follows a phase diagram that can be computed in advance. The theory predicts which task disappears first, at what fraction of failed agents, and whether the survivors split into a patchwork of single-task regions. All of this happens without a single message. The coordination medium is stigmergic, meaning agents coordinate only by modifying and sensing a shared environment, and fault tolerance becomes a predictable phase transition rather than a learned behavior.
Core Claim
We claim that coupling task fields through a competitive term in a shared environment field gives two things that independent pheromone fields and learned attention do not. The first is priority-ordered shedding: as agents fail, low-priority tasks collapse cleanly while high-priority ones keep near-full service, instead of every task degrading together. The second is a fault-tolerance ratio, the maximum fraction of failed agents a property survives, that can be predicted in closed form from mean-field theory. We then check that prediction against simulation, in the spirit of the formal ratios computed by Symbolic Synthesis of Fault-Tolerance Ratios in Parameterised Multi-Agent Systems. That work synthesized a 33% ratio for discrete Alpha Swarm Aggregation protocols but cannot handle continuous fields.
The environment maintains complex fields $\psi_k(x,t)$, one per task $k$. Each field represents the local "order" of task $k$, meaning how strongly that region is being served. The fields relax toward low values of the energy
$$F=\int \sum_k\Big[-a_k(x)|\psi_k|^2+\tfrac{b_k}{2}|\psi_k|^4+c|\nabla\psi_k|^2\Big]+g\sum_{k<l}|\psi_k|^2|\psi_l|^2\,dx.$$
The symbols mean the following. The drive $a_k(x)=w_k n(x)-\theta_k$ grows with the local density of healthy agents $n(x)$, scaled by a task priority $w_k$ and offset by a task cost $\theta_k$. The term $b_k$ caps the order's size. The term $c$ smooths it in space. The coupling $g$ sets how strongly tasks compete for agents.
Agents add to the drive of the task they perform, and each agent switches to the task with the largest local unmet drive $a_k-b_k|\psi_k|^2-g\sum_{l\neq k}|\psi_l|^2$. The swarm is therefore doing distributed gradient descent on $F$, using only local reads and writes.
Standard Ginzburg–Landau analysis then makes two predictions. When $g>\sqrt{b_kb_l}$, two orders cannot coexist uniformly and only one survives in each region. When only task $k$ is present and $n_0$ is the healthy density, task $k$ collapses at a fault fraction
$$f_k^*=1-\frac{\theta_k}{w_k n_0}.$$
This follows from setting $a_k=0$ at $n=n_0(1-f)$. If a stronger order $l$ is present, the competition term shifts this threshold upward by an amount set by $g|\psi_l|^2$.
Relation to Prior Work
The physics side tells us which behaviors to expect from competition. Under pressure, a native charge-density wave gives way to superconductivity in NbTe4, and pressure tuning of CsV3Sb5 produces a double superconducting dome that tracks its competition with charge order. Charge order disappears and reappears as composition changes in CuIr2Te4−xSex and in overdoped Bi2201. Exact lattice simulations in Competing Orders in a Nearly Antiferromagnetic Metal show a magnetic phase boundary bending back where superconductivity sets in. We borrow the phenomenology, not the microscopic physics. Suppressing one order frees capacity for another, and re-entrance (an order vanishing and then returning as a parameter changes) is possible. Both are testable predictions for swarms.
Disorder does not always destroy order. Multifaceted machine learning of competing orders in disordered interacting systems found a "charge-density-wave micro-emulsion", in which local droplets keep their order while global coherence is lost. We treat randomly placed crashes as disorder and ask whether the swarm forms analogous patches of single-task service.
The stigmergy side supplies the method and sets the bar. S-MADRL coordinates excavation robots through hand-designed pheromones that decay and diffuse, and finds emergent lanes and selective idleness. Its fields are single passive scalars with no competition between them, and it never injects crashes. CINOC controls classical physical fields, such as heat flow and fluid turbulence, with shared-parameter neural operators and handles actuator dropout zero-shot. In CINOC the field is the object being controlled; in our proposal the field is the coordination medium, and its dynamics are designed. Act2Comm shows that signaling through actions costs reward, roughly 20–30% at a bit error rate of 10⁻⁴ on Lucky Wheel. That cost motivates putting the signal in a persistent medium rather than in the task actions themselves.
On the fault side, AACFT isolates failed teammates with attention, which costs $O(M^2)$ in the number of agents $M$ and requires direct observation of teammates. The parameterised verification line, MCMAS-PFI and MCMAS-PFTS, gives exact fault ratios but only for discrete rule-based protocols.
Research Program
We will build the field medium as a simulator layer, a grid of $\psi_k$ values integrated with relaxational dynamics $\partial_t\psi_k=-\Gamma\,\delta F/\delta\psi_k^*$ plus agent deposits. It will plug into two existing environments.
The first is a multi-task extension of the S-MADRL excavation tunnel. Transport (high priority), bottleneck clearing (medium) and perimeter patrol (low) compete for the same robots. The second is the four Multi-Agent Particle Environment fault scenarios run through the FTMAL platform from AACFT, in which patrol and navigation become competing tasks.
Agent policies come in two versions. One is the analytic gradient rule above, which has no learning. The other is a learned, parameter-shared policy that observes only its local field patch, as CINOC does, so that cardinality and the number of failures do not change its input size.
Analysis proceeds in three steps. First, we derive the mean-field phase diagram over fault fraction $f$, priority ratio $w_1/w_2$, and coupling $g$. Second, we measure the empirical phase diagram through per-task service rates and a spatial structure factor of the task assignments, which tells us whether the swarm is uniformly mixed or broken into patches. Third, we compare predicted and measured collapse thresholds $f_k^*$.
Faults will include hard crashes, as in AACFT, and continuous drift, where an agent's speed or deposit strength degrades over time. Drift is a regime both the formal-verification and attention lines leave untreated. In this model, drift acts as a smooth reduction of $n(x)$, so the theory covers it without modification.
Baselines
The naive reference is uncoupled stigmergy: independent per-task pheromones with linear decay and diffusion, the S-MADRL style of field, combined with greedy task choice. It isolates whether the medium helps at all compared with independent learners, and S-MADRL itself reports that independent DQN falls to about 35 pellets per episode at 8 agents versus about 110 for S-MADRL.
The key ablation keeps the identical Ginzburg–Landau medium but sets $g=0$. This isolates the competition term, our central ingredient. A second variant adds hand-tuned per-task priority weights to the $g=0$ medium. It gets the same Bayesian-optimization budget, 200 trials, that we give to tuning $g$, $w_k$, and $\theta_k$.
The strongest learned competitor is AACFT, run through FTMAL on the same scenarios and fault schedules. It gets an equal environment-step budget and its released hyperparameters plus the same tuning budget. AACFT is allowed direct teammate observation, which our method lacks, so it is an advantaged comparison. The file gives no numerical post-fault AACFT result. The only reported number is vanilla MADDPG in Abandonment dropping from 87.2% to 38.2% completion after the fault, so we will rerun AACFT ourselves.
For reference ceilings we will include MAPPO with centralized training, which S-MADRL reports collapsing to near zero at 8 agents in the tunnel, and an oracle with explicit messaging, a centralized reallocator that knows every failure instantly.
Beating the field means three things. On high-priority service at 30% or more failed agents, the method must match or exceed AACFT and exceed the tuned $g=0$ variant by at least 10 points. On per-agent compute, it must scale linearly to 1,024 agents. Its predicted collapse thresholds must land within 5% fault fraction of the measured ones.
Why not just…?
Why not just use independent pheromones per task with priority weights? This is the real threat. If a tuned $g=0$ medium with priorities sheds tasks just as cleanly, then competition is unnecessary and the contribution shrinks to a mean-field prediction of fault ratios. The theory says uncoupled fields degrade every task together as $n$ falls, because each $a_k$ falls in proportion to $n$ with no exclusion. Coupled fields with $g>\sqrt{b_kb_l}$ instead hand the freed agents to the dominant task. The matched-budget ablation decides this directly, and if the two tie on shedding sharpness we will report that.
Why not just learn fault tolerance end-to-end, as AACFT does? Attention needs teammate observations that are absent when there is no communication channel, and it costs $O(M^2)$. It also gives no advance prediction of where collapse happens. A learned policy might still beat our analytic rule on raw reward at small $M$. For that reason we also test a learned, field-reading policy and report results across $M$ from 8 to 1,024.
Why not just send heartbeat messages and reallocate centrally? Where channels exist, that works, and the oracle shows how much is lost by going without messages. The proposal targets settings where messaging is jammed or forbidden, the regime motivating Act2Comm. A central reallocator is also a single point of failure, which we test by crashing the oracle's coordinator.
Proposed Experiments
The decisive check is quantitative agreement between predicted and measured collapse. We sweep the fault fraction from 0 to 0.8 in steps of 0.02 and plot measured per-task service against the predicted $f_k^*$, across three values of $g$ on either side of $\sqrt{b_kb_l}$. Seeing the predicted switch from coexistence to exclusion at that boundary would show that the phase theory actually governs swarm behavior.
The second check uses spatially clustered versus uniform crashes, the disorder strength $W$. We test whether a patchwork regime appears, in which each region serves one task but the global assignment is incoherent, analogous to the micro-emulsion of the disordered-torus study. We measure it with the structure factor and with a small classifier trained on field snapshots.
The third check looks for re-entrance. Does a low-priority task that disappears at moderate fault fractions return at higher ones, once the dominant task itself collapses? The theory predicts it does, and this is the most surprising and therefore most falsifiable prediction.
Throughput is reported against the S-MADRL tunnel and the FTMAL scenarios. We also report wall-clock per step against the linear-scaling claim and compare collapse thresholds with the discrete-protocol reference of 33% from MCMAS-PFTS on Alpha Swarm Aggregation. That last comparison is qualitative only, since the systems differ.
References
- Symbolic Synthesis of Fault-Tolerance Ratios in Parameterised Multi-Agent Systems
- Pressure induced superconductivity bordering a charge-density-wave state in NbTe4 with strong spin-orbit coupling
- Double Superconducting Dome and Triple Enhancement of <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mrow><mml:mi>T</mml:mi></mml:mrow><mml:mrow><mml:mi>c</mml:mi></mml:mrow></mml:msub></mml:math> in the Kagome Superconductor <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow><mml:mi>CsV</mml:mi></mml:mrow><mml:mrow><mml:mn>3</mml:mn></mml:mrow></mml:msub><mml:msub><mml:mrow><mml:mi>Sb</mml:mi></mml:mrow><mml:mrow><mml:mn>5</mml:mn></mml:mrow></mml:msub></mml:mrow></mml:math> under High Pressure
- Anomalous charge density wave state evolution and dome-like superconductivity in CuIr 2 Te 4− x Se x chalcogenides
- Re-entrant charge order in overdoped (Bi,Pb)2.12Sr1.88CuO6+δ outside the pseudogap regime
- Competing Orders in a Nearly Antiferromagnetic Metal
- Multifaceted machine learning of competing orders in disordered interacting systems
- Deep reinforcement learning for multi-agent coordination
- CINOC: Cardinality-Invariant Neural Operator Policies for Scalable PDE Control
- Actions Speak Louder Than Words: Rate-Reward Trade-off in Markov Decision Processes
- Towards Fault Tolerance in Multi-Agent Reinforcement Learning
- Verifying Fault-tolerance in Parameterised Multi-Agent Systems