Inspirations — physics & biology for the redesign
Cross-disciplinary mechanisms we're drawing on to break the two open walls — getting the team to divide the map (coordination / redundancy) and to stay linked at scale (the connectivity wall). Candidate ideas feeding the proposed Stack C redesign — not results.
The reframe that filters everything below: this is a small TEAM, not a swarm. At N ≈ 4–16 agents, large-N self-organization (ant colonies, starling murmurations, soap-foam coarsening, percolation in the thermodynamic limit) is a statistical effect that doesn't exist at our scale — at best a metaphor. And two obvious "swarm" fixes — a fixed 50/50 relay/explorer split and role rotation — were tried · ruled out. So the inspirations here are filtered for small-N, role-differentiated, position-determined mechanisms.
1 · The elastic-rope comms model
Replace our current link model — a binary, memoryless cutoff at comm-range (link at d=4.9, gone at d=5.1, no warning, no state) — with an elastic rope between agents that builds tension as they separate and snaps under enough strain. It maps cleanly onto a material's stress–strain curve:
| Rope physics | Comms meaning | Per-edge variable |
|---|
| Elastic (stretches, springs back) | full link — within nominal range, healthy comms | tension τ = k·max(0, d−d_rest) |
| Yield / plastic (deforms, accrues strain) | degraded link in a stretch band — alive but lower bandwidth, rising staleness | strain ∫(τ−τ_yield)⁺ dt |
| Fracture (snaps) | link lost | break when d > r_break or strain > budget |
What it buys (the genuinely new parts):
- The elastic band = free temporary over-range = grace / intermittent connectivity. Agents can briefly leave contact to cover ground as long as the rope doesn't break — so they spread to the elastic limit, not the range limit. This is the main untried lever for the connectivity wall.
- Strain-then-snap = Age-of-Information with a budget — and a covert-attack surface: a compromised agent can manipulate tension without ever snapping a rope (free-ride by staying central, or fatigue a critical link to near-breaking). Tension is local and oracle-free (unlike the Fiedler value), so it's deployable.
- Per-edge state
(tension, strain, health) = relational edge features for the graph belief.
Don't misread the rope. The restoring-force reading (a spring that pulls everyone back together) is just the soft tether we already tried — it loses to the hard guardrail and, at the limit, gives the clump. Tension must be a constraint (don't exceed the break budget), not a reward term (a "minimize tension" objective Goodharts into the huddle, as adaptive-λ did). Ropes attach on contact, not as a complete mesh. And the rope is a connectivity/comms idea — it does not fix coverage division.
2 · Physics — useful as theory, not as large-N dynamics
Physics helps two ways: mechanism metaphors (how an agent acts) and theory (what's achievable). For us the theory is the bigger prize — it can turn one of our "failures" into a finding. But most large-N physics is statistical, so it's descriptive at our scale, not a control law:
| Inspiration | What it gives us | Scale fit |
|---|
| Percolation / random geometric graphs | The connectivity wall is a phase transition: below a critical density the giant component shatters. Reframes "we couldn't stay connected at 32²" as "physics says one rigid component is improbable here → use grace," not a policy bug. | large-N · descriptive |
| Soap foam / surface tension (Plateau) → Voronoi/CVT | Bubbles tile space into disjoint cells = a spatial partition = the physical face of CVT/LPAC = anti-redundancy. | CVT target fits small-N |
| Reaction–diffusion (Turing) / stigmergy | Local inhibition → even spacing. A decaying "I claimed here" field = decentralized anti-redundancy (and the decay is AoI again). | pattern is large-N |
| Self-organized criticality (sandpiles) | A team at the edge of connectivity shows avalanches of all sizes from tiny perturbations — the micro→macro amplification thesis in physics; predicts attack leverage peaks at criticality. | conceptual |
| Spider web / tensegrity / granular force-chains | Pre-stressed, damage-tolerant networks = r-robustness (redundant load paths > mere λ₂-connectivity); force-chains = which agents are load-bearing = the amplification map. | fits small-N |
The pattern: constraint- and field-based physics helps us; force-based physics (charges, springs-as-attraction, boids) is just soft potential fields — which we've shown lose to hard constraints and deadlock in local minima. Highest-value pulls: percolation as the theory of the wall, and self-organized criticality as the theory of attack leverage.
3 · Nature at the right scale — small-group cooperative hunting
Small predator groups (3–12) solve role division the way we need to — and it's the opposite of a 50/50 rotating split:
- Lions hunt in groups of 3–7 split into "wings" and "centers": wings circle out and drive the prey; the center (the heaviest, most experienced individual) waits in ambush. [overview]
- Dolphins (Cedar Key): one "driver" herds fish against a "barrier" line of the others — 1-to-many. [pack hunting]
- Wolves: role by individual attribute — light fast females herd; heavy males take down.
Three lessons, all against what we tried:
- Ratio is asymmetric and emergent, never a fixed 50/50 (several wings, one center; one driver, many barriers).
- Roles are stable within a task, not rotated — the center stays the center. (Fairness is handled across many hunts by reciprocity, not mid-task rotation.)
- Role = a function of position — wing vs. center is set by where you are relative to the prey and the team, not by a quota or a clock.
4 · The small-N toolkit that actually fits
Three rigorous, citable tools that are native to few agents (and that we have not tried):
- Rigidity & persistence theory — the rope made formal. "Rigidity plays the role of connectivity when only distance (range) measurements are available." A graph can be connected but floppy; rigidity means the team's shape is recoverable from inter-agent distances, it has a rigidity eigenvalue (richer than the Fiedler value), and it's maintainable decentrally from range only while the topology changes freely. It tells you which links are load-bearing → who must relay, and degrades gracefully. [Zelazo et al., RSS] [IJRR] [persistence]
- Shapley value / marginal contribution — the army-ant cost–benefit rule, formalized: value each agent by its average marginal contribution. Picks relays (highest marginal connectivity contribution) and gives a credit-assignment signal for training; tractable at small N (exponential in N — fine at ≤16, hopeless at 1000, another reason swarm methods don't fit). [SHAQ: Shapley-value MARL] [Shapley MRTA]
- Communication-constrained multi-robot exploration — the directly-relevant subfield (search the small-team framing, not "swarm"). Its standard structure is exactly explorer / relay / relay-dispatcher roles — and they're asymmetric, never an even split. [Amigoni/Banfi/Basilico survey] [MOCHA] [SLEI3D 2026]
5 · The synthesis — the relay/explorer "role law"
Physics, biology, and small-team robotics converge on the same control law, and it is the opposite of a symmetric rotating split:
- Role = f(structural position) — relay vs. explorer is determined by whether you're load-bearing in the graph right now (rigidity / Shapley-on-connectivity), the way a lion's role is set by position — not a quota, not a turn.
- Ratio emerges asymmetric — often few relays / many explorers (or the reverse near a chokepoint). Forcing 50/50 imposed a symmetry the geometry rarely wants.
- Pick relays by marginal contribution (Shapley), not by clock — which doubles as the training credit signal.
- Maintain rigidity, not just connectivity — the right structural target for a distance-constrained (roped) small team, deployable from range alone.
Honest caveat: this gives the who-holds-the-net law — it does not by itself fix coverage division (the redundancy / flooding problem). That still needs the goal/region action head (so an agent can actually go hold the bridge its structural position assigns it) plus a partition target (foam/Voronoi/CVT). Rigidity + Shapley give the role law; the action-representation change gives the ability to execute it. See Architecture → Stack C.
6 · Already tried / ruled out
So we don't re-litigate — these are closed unless reopened with the stated reason addressed:
| Approach | Verdict |
|---|
| Fixed 50/50 relay/explorer ratio | imposes a symmetry nature avoids; ratio should emerge from position/need |
| Role rotation / turn-taking | fights the natural stable-complementary-role structure within a task |
| Soft tether / cohesion forces (potential fields, boids) | lost to the hard guardrail by ~20 pts; deadlocks in local minima |
| Adaptive-λ (Lagrangian on connectivity) | degenerates into the all-relay clump (λ→1.5, coverage dies) |
| Fiedler oracle as a feature | null — a perfect global signal didn't help; the limit is the action space |
| Per-step 1-step move head | the action-representation bottleneck; can't represent "claim a region" |
Status. These are candidate inspirations, grounded in the cited literature, feeding the
Stack C redesign (proposed, not yet approved). They attack the
measured walls and the
failure modes, and the rope/AoI & force-chain ideas double as the covert-attack surface for the
resilience question. Our prior literature sweeps were arXiv-weighted, so the rigidity, Shapley-MARL, and cooperative-biology threads still owe a targeted robotics-/biology-venue review before any paper claim.