worldmodel-gym

WorldModel Gym is a reproducible long-horizon planning benchmark + evaluation platform for imagination-based agents.

Quickstart (30 seconds)

make setup
make demo

make demo will:

• start the API + web stack with Docker when available
• fall back to local API execution when Docker daemon is unavailable
• run one benchmark evaluation
• upload artifacts and populate leaderboard data

Open:

• http://localhost:3000 (web dashboard)
• http://localhost:8000/docs (FastAPI docs)

Run a single evaluation

.venv/bin/python -m worldmodel_gym.eval.run \
  --agent random \
  --env memory_maze \
  --track test \
  --seeds 211,223 \
  --max-episodes 2

Artifacts are written to runs/<run_id>/:

• metrics.json
• trace.jsonl
• config.yaml

Monorepo layout

• core/: environments, traces, eval harness
• planners/: MCTS, MPC-CEM, trajectory sampling
• worldmodels/: deterministic/stochastic/ensemble latent models
• agents/: baseline agents and registry
• server/: FastAPI leaderboard + run artifact service
• web/: Next.js dashboard
• mobile/: Expo viewer
• paper/: draft PDF + LaTeX sources

Dev targets

make lint
make test
make paper
make deploy
make stop
make deploy-public
make stop-public
make deploy-vercel

Free Cloud Deploy

• API: deploy render.yaml on Render Blueprint (free web service).
• Web: deploy web/ on Vercel Hobby with NEXT_PUBLIC_API_BASE set to the Render API URL.
• Full steps: docs/DEPLOYMENT.md.