Skip to content

About

Connectome-constrained agents learning to fight through reinforcement learning in FightingICE.

Resources

Stars

3 stars

Watchers

0 watching

Forks

Latest commit

 

History

802 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Connectome Fighter

Connectome Fighter is a cloud research testbed that connects the MaleCNS v1.0 Drosophila connectome to FightingICE through a pinned implementation of the Shiu et al. leaky integrate-and-fire dynamics. The project records real MaleCNS body-ID activity, keeps game I/O mappings explicit, and separates biological structure from project-defined reinforcement, plasticity, and embodiment assumptions.

Production viewer: https://connectome-fighter.vercel.app/ · Dedicated LIVE: deployment target is configured through the repository variable CONNECTOME_PUBLIC_BASE_URL; no unrelated Vercel project is a valid Connectome target. · Research ledger · Status · Roadmap · Public surface contract

Canonical control path

FightingICE numeric observation
        ↓
versioned project-defined Poisson sensory interface
        ↓
MaleCNS v1.0 real connectivity
        ↓
pinned Shiu et al. LIF dynamics
        ↓
real output-body spike counts
        ↓
versioned action groups
        ↓
FightingICE action

The canonical path does not replace the fly substrate with an MLP, RNN, PPO policy network, or the repository's older custom sigmoid scaffold.

Physical FlyBody spectator

The physical fly is the upstream TuragaLab/flybody MuJoCo body pinned at commit d015e9bfe441bd90ae431bac24c55cb74bdbce26.

annotated MaleCNS motor / descending activity
        ↓
project-defined bounded adapter
        ↓
59 FlyBody actuators
        ↓
MuJoCo physics
        ↓
FlyBody render

The adapter is malecns-annotated-motor-to-flybody-tripod-v2. It is a project interface, not a claimed biological motor-neuron→muscle map. FightingICE x/y/action values do not position the FlyBody body. Zero neural motor drive produces no active gait. FlyBody, FightingICE ScreenData, and neural anatomy are spectator outputs and never policy inputs.

Scientific boundaries

  • MaleCNS anatomy is biological data.
  • Shiu LIF dynamics are a published neural-dynamics model applied to that anatomy.
  • FightingICE feature-to-sensory and output-to-action mappings are project-defined interfaces.
  • MaleCNS→FlyBody actuator mapping is project-defined.
  • Reward and plasticity rules are research-added assumptions and remain versioned.
  • Neural activity is not treated as proof of causal biological function; causal claims require interventions or ablations.

Current demonstrated state

  • MaleCNS v1.0 provenance/import: PASS
  • pinned Shiu LIF reference/runtime: PASS
  • strict runtime: approximately 156,675 neurons / 6,025,920 recurrent synapses under the project filter
  • real MaleCNS-controlled FightingICE rounds: PASS
  • body-ID spike/event logs: PASS
  • compilerless Brian2 Cython runtime bundle: PASS
  • FlyBody real MuJoCo physics contract: PASS in independent CI
  • FlyBody action dimension: 59
  • neural-drive-vs-zero trajectory divergence: PASS
  • dedicated Next.js public control/viewer implementation: IMPLEMENTED; production deployment verification pending
  • dedicated Vercel shared-LIVE E2E: NOT YET VERIFIED
  • GARNET generation 2: approved read-only inference state
  • ZEN canonical baseline: current P2 serving contract
  • continuous production reward-driven learning: OFF

The prior unrelated Unjuno/live application is not a Connectome deployment target and is not modified by the dedicated workflows.

Dedicated public architecture

GitHub rolling runtime release
        ↓
SHA-addressed persistent runtime base
  (runtime + FightingICE fonts + OSMesa)
        ↓
fixed-name connectome-fighter-live-broadcast Sandbox
        ↓
FightingICE + two MaleCNS/Shiu workers
        ↓
Official ScreenData + neural activity + FlyBody physics
        ↓
one shared public viewer surface

The dedicated app lives in this repository under app/ and lib/. Its control API is /api/live; runtime materialization is /api/runtime-base. It reads the Vercel deployment's own VERCEL_PROJECT_ID and contains no hard-coded identifier for another Vercel project.

Production smoke workflows use CONNECTOME_PUBLIC_BASE_URL. If that variable is unset, they skip rather than falling back to another site.

Runtime publication boundary

Runtime publication is GitHub-only:

publish-arena-runtime-bundle → precompile-arena-brian2-cython-cache → publish-flybody-runtime-addon.

These workflows publish immutable release assets but do not stage them into a Vercel project. The dedicated Vercel control plane owns runtime-base staging. This prevents release jobs from mutating an unrelated deployment.

Learning status

Continuous canonical learning is not production-active. Candidate training and append-only lineage history remain on GitHub and cannot auto-promote into the Vercel inference state.

License

Project-authored source code is MIT licensed. FightingICE, MaleCNS/connectome data, FlyBody, papers, and other third-party resources retain their own licenses and terms; see THIRD_PARTY.md.

About

Connectome-constrained agents learning to fight through reinforcement learning in FightingICE.

Resources

Stars

3 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages