Yann LeCun — former Meta chief AI scientist and co-founder of world model startup AMI Labs — has teamed up with Mila, Brown University, Oxford University, and LanceDB to release a new open-source platform called stable-worldmodel (swm for short). The project is led by Lucas Maes, a PhD student at Mila, with AMI Labs researcher Randall Balestriero as corresponding author. The goal: standardize and reproduce world model research in embodied AI.
World model research is currently a mess. The field suffers from three major bottlenecks: fragmented codebases, painfully slow video data loading, and a lack of standardized generalization benchmarks. Different teams often have to build their own data pipelines and core algorithms from scratch, leading to duplicated effort, inefficiency, and subtle inconsistencies that undermine fair comparisons and reproducibility.
To tackle the I/O and video loading problem, stable-worldmodel introduces a high-performance data layer based on the Lance format, along with conversion tools for MP4 video, HDF5, and LeRobot robot datasets. On the algorithm side, the platform integrates a bunch of mainstream world model baselines — including DINO-WM, PLDM, LeWM, GCBC, GCIVL, and GCIQL — and provides planning solvers like CEM, MPPI, and gradient-based methods.
The evaluation environment supports customizable changes to visual, geometric, and physical factors, letting researchers systematically test a model’s dynamics understanding, control performance, representation quality, and out-of-distribution generalization. By bundling a complete data pipeline and evaluation framework into a single platform, stable-worldmodel aims to slash the R&D overhead and infrastructure costs for world models, offering a highly reproducible standardized base for embodied AI research.