How galilai-group/stable-worldmodel works
stable-worldmodel is a research platform for standardized world-model data collection, training, and model-predictive-control evaluation. The documented workflow configures a vectorized World over supported environments, records episodes through the dataset-format interface, loads clips for model training, then evaluates a policy that plans with a model and solver. Groups distinguish environments, orchestration, data, learning, and planning. The README documents additional dataset backends and reference baselines; those are represented without asserting unsampled implementation wiring.
Main entry points: world.py World · formats Dataset formats · env_pool.py Environment pool · format.py Format registry
Components 15
Environments
- Environment suites
stable_worldmodel/envs
World orchestration
- World
stable_worldmodel/world/world.py - Environment pool
stable_worldmodel/world/env_pool.py - Environment wrappers
stable_worldmodel/wrapper/default.py
Dataset pipeline
- Format registry
stable_worldmodel/data/format.py - Dataset formats
stable_worldmodel/data/formats - Lance backend
stable_worldmodel/data/formats/lance.py - Dataset abstractions
stable_worldmodel/data/dataset.py
World models
- Training recipes
scripts/train - World-model baselines
stable_worldmodel/wm
Planning and evaluation
- World-model policy
stable_worldmodel/policy.py - Planning solvers
stable_worldmodel/planning/solver - Evaluation
stable_worldmodel/planning/evaluator.py - Episode results
stable_worldmodel/world/world.py
Other components
- Researcher
Connections 18
- Researcher → World: configures
- World → Environment pool: steps environments
- Environment pool → Environment wrappers: processes observations
- Environment pool → Environment suites: dispatches to
- World → Format registry: records episodes
- Format registry → Dataset formats: selects writer
- Format registry → Dataset formats: selects reader
- Dataset formats → Lance backend: includes backend
- Researcher → Dataset abstractions: loads dataset
- Dataset formats → Dataset abstractions: provides samples
- Dataset abstractions → Training recipes: supplies clips
- Training recipes → World-model baselines: trains baselines
- Researcher → World-model policy: configures policy
- World-model policy → Planning solvers: requests plans
- Planning solvers → World-model baselines: predicts with model
- World → World-model policy: runs policy
- World → Evaluation: evaluates episodes
- Evaluation → Episode results: produces metrics