What is zr-0?

ZR-0 is an AI model built to control physical robot arms and other robots, letting them look at a camera feed, understand a spoken or written task, and figure out what physical movements to make. This kind of model is called a Vision-Language-Action model, meaning it takes in images and language and outputs actions rather than just text. The model has 2.6 billion parameters and combines two parts working together. One part is a vision and language model that reasons about the scene and the task in words, similar to how a person might narrate their own plan step by step before acting. The other part is a separate action generating component that turns that reasoning into smooth, continuous movements for the robot. The two parts are trained together so the reasoning step can be skipped once the model is running, which lets it decide on an action roughly every 90 milliseconds on suitable hardware. ZR-0 was trained on a large dataset of more than 400,000 robot demonstration recordings across different robot types, including robotic arms, a humanoid platform, and dual arm setups, most of which include detailed written explanations of what the robot is doing and why at each step. After this initial training, the README reports that fine-tuned versions of the model perform well on several published robotics benchmarks covering single arm tasks, humanoid tabletop tasks, and two armed manipulation tasks, as well as tests on a real physical robot arm. To use it, you clone the repository, set up a Python environment with Conda, and install a specific set of dependencies including a library called flash attention that needs to match your GPU, Python, and PyTorch versions exactly. Pretrained model checkpoints, along with several versions already fine-tuned for specific benchmark datasets, are hosted separately and linked from the README. Evaluation works by running the model as a server that a robot or simulated environment connects to over a network connection, sending it images and task descriptions and receiving back a short sequence of actions to execute.