What is starwam?

StarWAM is a research codebase for building what its authors call World-Action Models, a type of robot control system that combines an AI video generation model with a module that predicts physical actions. The idea is that a model which has learned to generate realistic video of the world can also be adapted to help a robot decide what to do next, since predicting what happens visually and predicting what action to take are related problems. This is early stage research code, and the authors say more model variants, benchmarks, and pretrained models will be added over time. The codebase is built to be modular, meaning researchers can mix and match different existing video generation backbones, such as Wan2.2 or Cosmos-Predict2, with different approaches to combining video and action prediction. It currently supports three different families of these combined models, each with a different internal design for how video information and action information interact inside the network. Configuration for experiments is handled through YAML recipe files that get loaded into typed Python objects, rather than requiring a separate configuration framework. The project includes example setups for two established robotics benchmarks, LIBERO and RoboTwin 2.0, with instructions for training and evaluating models on each. Pretrained model checkpoints are hosted separately on ModelScope and can be downloaded with a small command line tool, including specific weights for both of the supported model families on the LIBERO benchmark and a dual-arm setup for RoboTwin 2.0. The project acknowledges that it draws heavily on ideas and code from several other open research projects in robot learning and video generation, crediting them by name. The readme includes a roadmap of completed and planned work, including a technical report that has not yet been released, and a citation section that is still pending a formal reference. The readme does not state a specific software license for the code.