stable-worldmodel
Python toolkit for collecting data, training, and evaluating world models
Featured by Willet
What it is
stable-worldmodel gives one interface for gathering trajectory datasets across standardized simulated environments, training world models on them, and scoring them with model-predictive control. It includes planning solvers such as CEM, reference baselines like LeWM and a DINO-WM reproduction, and several dataset formats (Lance, HDF5, video, LeRobot) with conversion between them. Machine learning researchers working on world models and planning use it to keep experiments reproducible and comparable.