An index of where each part of the project now lives. The original markdown corpus has been folded into the pages of this site; the source binaries (PDFs, LaTeX, the taxonomy YAML tree) are preserved as downloads below.
assets/sources/. Section 3 notes the code repositories that are not bundled here.Each topic below has been written up as a page. Follow the link for the full treatment.
The MARL / agentic-AI field map (2019–2026): the catalog of entries and surveys organised by interaction mode, the cooperative-vs-collaborative axis, and the candidate-gap analysis that positions this work. Auto-generated from the taxonomy YAML (downloadable in Section 2).
The formal model: the Networked Dec-POMDP / POSG tuple with a dynamic comm graph, the CTDE state split, the attack-channel and stealth/intervention/detectability parameters. Mirrors formalism.tex (downloadable in Section 2).
The Sense / Organize / Act mission taxonomy: the ternary-simplex placement of deployments and hardware anchors (Kilobot, CADRE, OFFSET, Starling), and why coupling mechanism determines the attack surface. Built from the standalone mission-taxonomy LaTeX (downloadable in Section 2).
The related-work review: prior swarm/MARL taxonomies, the multi-robot canon worth borrowing (connectivity estimation, coverage control, connectivity-constrained exploration), and the conjunction of axes that remains novel.
The quantitative results from RedWithinBlue: the cooperative coverage ladder, the n=60 adversarial-validation sweep (the action-stream channel is statistically silent; damage scales with how many agents are compromised; the quietest "stay" attacker is most damaging; co-evolution is a defender-side fragility).
The JAX/Flax pipeline: the PureJaxRL single-JIT train loop, the CTDE actor / central-critic design, the warm-start weight-transfer ladder, and the environment / comm-graph API.
The emergent-behaviour story: explorer/relay role differentiation, symmetry-breaking, the connectivity guardrail, and the graph-belief swarm design (Compass / SLAM / Emergence).
The research narrative: how the question evolved from single-agent coverage to covert-misbehavior resilience, the negative results (TD(0) collapse, statistically silent belief corruption), and the decisions taken and abandoned.
The NeurIPS-style write-up "Stealth Attacks on Swarms": threat model, methods, and the ~45-figure results section.
The original source files, preserved verbatim under assets/sources/. These are the binaries the pages above were built from.
The standalone formal model — the compiled formalism document (Networked Dec-POMDP, attack channels, stealth parameters).
The swarm-resiliency write-up (v1.3) — early framing of covert-misbehavior propagation and macro mission failure.
The mission-family reference: the deployment landscape behind the mission taxonomy.
The compiled "Stealth Attacks on Swarms" paper PDF (see Paper for the in-site version).
The full Overleaf project archive for the paper — LaTeX sources, sections, figures, and bibliography.
The adversarial-ML report — background survey of attacks and defenses informing the threat model.
The ExoRL report — exploratory reinforcement-learning notes feeding the early method choices.
The compiled standalone mission-taxonomy document, including the ternary-simplex figure (see Missions for the in-site version).
LaTeX source for the mission-taxonomy document above (its missions.bib and Images/ live alongside it in the same folder).
The auto-generation source for the Field Map: the generated README.md plus the data/ entry/survey YAML, schema/ pydantic models, and build Makefile. Browse the folder from the linked README.
RedWithinBlue/ (the JAX MARL experiments engine), zymera_env/ (the swarm-explore engine), and zymera_lab/ (the consolidated simulator + lab) are maintained in their own git repositories and are not bundled into this static site. The narrative, architecture, and results drawn from them are written up in the pages linked in Section 1.