FineART: Scale AI’s First Open-Source Robotics Dataset and Model

Scale AI and Hugging Face have released FineART, the largest open robotics dataset with detailed subtask annotations. The dataset contains over 40,000 episodes and 533,000 labeled steps, focusing on complex manipulation and long-tail tasks. Accompanying the data is FineART-VLA, a vision-language-action (VLA) model that plans subtasks before executing actions, improving performance on long-horizon tasks. This model architecture allows it to better follow spatial instructions and generalize across different robot hardware. The dataset and model are integrated into the open-source LeRobot ecosystem to accelerate research in physical AI.