wtf is opendw?

DW05 is an open source AI model for robotics research, built to predict what happens next in a robot's environment based on the actions it takes. The project describes it as a world model, meaning a system trained to imagine future video frames given a starting image and a planned action, while also generating the actions themselves and estimating how good a given state is likely to be. These three pieces, predicting future video, generating actions, and estimating value, are combined inside one framework rather than kept as three separate models. Under the hood, DW05 is built from a shared backbone, called Wan in the README, that branches into three specialized expert components: one for video, one for action, and one for value estimation. The value estimation part is described as not yet finished and will be updated in a future release. The project ships pretrained model weights on Hugging Face in two versions, a base model trained on a mix of data sources, and a version fine tuned on a benchmark called Robotwin 2.0 for evaluation purposes. Getting it running requires Python 3.10 or newer plus a CUDA enabled build of PyTorch, since this is a GPU heavy research model rather than something that runs on an ordinary laptop. Training the model expects data in a specific format: episodes recorded from robot cameras, saved as JSONL files with one line per frame, each including camera images or video clips, the robot's state and planned actions, a text description of the task, and cached text embeddings. The README documents this format in close detail and includes a small demo dataset split so people can see the expected folder layout before attempting to use their own robot data. The repo also includes example scripts for running inference, including a demo that shows the model predicting a future video conditioned on a robot's observations and planned actions. This is aimed squarely at robotics and embodied AI researchers rather than general developers, and it is released under the Apache 2.0 license.