Stable WorldModel
github.com/galilai-group/stable-worldmodel- Category
- AI Tools
- Rank
- No. 1647Tools index
Previous survey · No. 1548 ·
- Type
- APP
- Builder
- galilai-group
- GitHub
- 2.2k stars
- Latest release
- 0.1.1
- Date
About
A platform for reproducible world-model research and evaluation.
What it does
Stable WorldModel connects environment rollouts, episode storage, model training inputs, and model-predictive control. Researchers attach a policy, collect trajectories across vectorized Gymnasium environments, load them through a format registry, then evaluate a supplied world model with planning solvers.
Why it's ranked here
The case is strong because the project tackles repeated experimental plumbing, not merely model code. It standardizes collection and evaluation, exposes controlled visual and physical variation, includes several planning methods, and documents reproducible dataset inspection and conversion. Its active-development warning keeps the verdict qualified.
What's good
The data layer supports Lance, HDF5, folders, video, and read-only LeRobot datasets behind one registry. Writers append by default, readers detect formats, and conversion preserves an episode-oriented workflow. The command line can inspect, preview, merge, and convert datasets without requiring Python code.
Tradeoffs
APIs may change between minor versions. The base installation omits training, environment, and additional format dependencies. LeRobot support is read-only and requires Python 3.12 or newer. Planning with observation history pads early episodes with repeated first frames and zero actions, creating synthetic context that differs from training clips.
How to use it well
Use it for world-model experiments that compare models, planners, storage formats, or robustness under controlled environment shifts. It earns a place between Gymnasium environments and PyTorch training code. It does not replace the model itself: researchers must train or supply an object that computes planning costs.
Technical notes+
pyproject.toml defines a Python 3.10+ package built with uv_build, a swm console entry point, core Torch and Gymnasium dependencies, and separate train, env, format, and lerobot extras. docs/quick_start.md describes the vectorized World abstraction, pluggable dataset registry, PyTorch-compatible loading, and model-predictive evaluation. docs/cli.md covers dataset listing, inspection, preview, conversion, merging, environment variation inspection, and checkpoint listing. docs/api/solver.md documents CEM, iCEM, MPPI, predictive sampling, gradient, projected-gradient, categorical CEM, and augmented-Lagrangian planners. docs/api/buffer.md explicitly describes synthetic warm-up history.
Observed
- License
- MIT
- Primary language
- Python
- Python requirement
- Python 3.10 or newer; the LeRobot extra requires Python 3.12 or newer
- Packaging
- Published as stable-worldmodel for pip or uv installation, with optional train, env, format, lerobot, and all extras
- Interfaces
- Python library and swm command-line interface
- Dataset formats
- Lance, HDF5, folder, video, and read-only LeRobot adapters
- Environment interface
- Included environments follow Gymnasium, and additional environments can conform to that interface
- Build backend
- uv_build
Read from README.md, pyproject.toml, docs/cli.md, docs/index.md, docs/baselines.md, docs/quick_start.md, docs/envs/ale.md, docs/envs/dmc.md, docs/envs/ogb.md, docs/api/world.md, docs/api/buffer.md, docs/api/policy.md, docs/api/solver.md.
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.