galilai-group/stable-worldmodel

What it solves

It provides a unified, reproducible platform for world model research, streamlining the three main stages of the workflow: collecting data, training models, and evaluating them using model-predictive control (MPC). It removes the need for researchers to write boilerplate code for data handling and planning solvers, allowing them to focus on the model architecture and objective functions.

How it works

The platform integrates several key components into a single interface:

• Data Management: A format registry supports multiple backends (LanceDB, HDF5, folder, video, and LeRobot) for recording and loading datasets, with built-in tools for one-shot migration between formats.
• Environment Suite: It provides a large collection of standardized environments from sources like DeepMind Control Suite, Gymnasium, OGBench, and Atari, many of which include "factors of variation" to test how models generalize to different visual or physical settings.
• Planning Solvers: It ships with reference implementations of sampling-based solvers (e.g., CEM, MPPI) and gradient-based solvers (e.g., SGD, Adam) to translate world model predictions into actions.
• Unified API: A simple Python API allows users to collect data with an expert policy, train a model, and evaluate it using a WorldModelPolicy paired with a solver.

Who it’s for

AI researchers and engineers working on world models, embodied intelligence, and model-predictive control who need a standardized way to benchmark and reproduce their results across diverse environments.

Highlights

• Extensive Environment Library: Includes a wide range of tasks from simple control (CartPole) to complex robotics (Fetch) and Atari games.
• High-Performance Data I/O: Optimized for fast indexed reads and append-friendly storage using LanceDB.
• Built-in Baselines: Includes reference implementations for JEPA-based models (LeWM, DINO-WM) and other behavioral cloning or RL baselines.
• Generalization Testing: Integrated factors of variation make it easy to evaluate zero-shot generalization to distribution shifts.