GPS-denied autonomy OS for unmanned platforms -- stereo VSLAM, edge targeting, 3D battlefield reconstruction, offline swarm coordination.

cd perception
cp .env.example .env        # set VIDEO_SOURCE, WS_HOST, WS_PORT
pip install -r requirements.txt
python main.py

Jetson Nano -- first-time setup:

python export_trt.py        # exports yolo11n.engine (run once)
# then in .env: YOLO_MODEL=yolo11n.engine  DEVICE=0
python main.py

See perception/README.md for full docs.

React + TypeScript app in frontend/, managed with Bun.

cd frontend
bun install
bun dev

Other commands:

bun run build    # production build → frontend/dist
bun run preview  # preview production build
bun run lint

From the repo root, start the FastAPI training/sim server on port 8787:

uv run --project swarm uvicorn swarm.backend:app --host 127.0.0.1 --port 8787

The process stays idle until the frontend calls POST /api/train/start or sim endpoints. Requires uv and deps in swarm/ (uv sync --project swarm if imports fail).

If port 8787 is already in use:

lsof -ti :8787 | xargs kill

Run the backend (above), then the frontend. Clicking Train Policy starts training only for the selected gym environment.

In a second terminal:

cd frontend
bun install
bun dev

Enter a gym environment, tune battlefield params, and click Train Policy. Metrics stream over WebSocket; on completion the service exports frontend/public/policies/<env_id>/policy.onnx and unlocks Mission Sim.

Optional one-terminal dev mode:

VITE_AUTO_TRAIN_SERVICE=1 bun dev

Optional env overrides:

• VITE_TRAIN_API_URL — default uses the Vite /api proxy to http://127.0.0.1:8787
• VITE_TRAIN_WS_URL — default ws://127.0.0.1:8787/ws
• VITE_TRAIN_TIMESTEPS — default 12000
• VITE_ALLOW_HEURISTIC_MISSION=1 — skip policy gate in dev