1Defensible contribution
We study how covert, within-bounds misbehavior by a single agent propagates to mission-level failure in a learned, decentralized, connectivity-aware coverage swarm — and how that propagation is governed by the agent's role and graph-position. The substrate is a size-invariant graph belief that doubles as the attacker's surface and the team's stealth detector; resilience is posed as a stealth–damage frontier rather than a single attack. This bridges three literatures that do not currently meet: connectivity-aware role allocation (homogeneous, no covert adversary), covert c-MARL poisoning (game benchmarks, trigger-based, no spatial/role structure), and resilient consensus / cascade modeling (detectable or non-adversarial). These objects are made precise in the project's formalism (formalism.tex): the stealth budget as KL-undetectability, the break budget, the propagation model, and the brittleness frontier — see the theory page.
2Limitations & threats to validity
- Compute. CPU-only; the host-side engine is fast at 16×16/4 but the MARL-at-32×32 path is slow, so several scale claims are zero-shot / few-shot transfer rather than full retrains.
- Literature-review tooling. The connectivity review completed cleanly (96→24). The covert-flank review's automated synthesis degraded (transient structured-output failures + an API overload), so that flank was re-established by direct search — treat it as direct-read, not PRISMA-verified.
- Search coverage. Both reviews were arXiv-weighted; IEEE / ACM / Springer robotics venues (ICRA, IROS, RSS, CoRL, T-RO) are under-represented. A targeted robotics-venue sweep is the outstanding check before locking any novelty claim.
- Weakest novelty leg. The covert single-agent → team-failure framing is partly taken (c-MARL backdoors); the defense leans on the spatial + role/position + belief + stealth-frontier conjunction, never the bare framing.
3Open next steps
- Targeted robotics-venue literature sweep + a clean re-run of the covert systematic review.
- Connectivity as a λ₂ constraint (computed from the belief) and an r-robust topology objective for the adversarial setting.
- Intent-signalling role negotiation (no consensus) wired into the ES switcher.
- Inject covert misbehavior at the belief / relay / intent channels and measure the stealth–damage frontier.
4Resources — the full program
- Theory: formalism.tex (workspace root) — the mission-centered resiliency formalism.
- Engine + code: zymera_env/swarm_explore/ — host-side JAX engine (compass, ES role-switcher, GCRN belief, belief-wiring); RedWithinBlue/ — the MARL experiments engine.
- Experiment artifacts: zymera_env/experiments/swarm_explore/ — checkpoints, metrics, GIFs, per-run reports.
- The paper: report/ — Stealth Attacks on Swarms (NeurIPS-style results report).
- Field maps: marl_taxonomy/ (70-entry interaction-mode field map + gap reports) · docs/mission-taxonomy/ + docs/taxonomy-refs/.
- Drafts & decks: references/ (Swarm Resiliency, mission family, advML, ExoRL) · presentations/ (Red within Blue, Swarm Resiliency / Boeing PI).
- Preserved source documents (in-site): the project's source documents are now mirrored in-site under assets/sources/ and indexed on the Docs page — the formalism (formalism.tex / formalism.pdf), the Stealth Attacks on Swarms paper plus its Overleaf source archive, the Swarm Resiliency and Mission Families drafts, the advML and ExoRL reports, and the full marl_taxonomy YAML.