Scale AI and Hugging Face open a 1,718-hour robot dataset
FineART pairs more than 40,000 bimanual demonstrations with dense step-by-step labels and a model that predicts its next subtask.
Scale AI and Hugging Face have released FineART, an open dataset for training robots to perform multi-step physical tasks. It contains 40,543 teleoperated demonstrations totaling about 1,718 hours on a two-armed ALOHA-style robot.
The collection covers 151 tasks and includes 533,913 human-written subtask labels. Unlike datasets that give only one instruction for an entire episode, FineART marks a sequence such as identifying an object, grasping it and placing it in the correct location. The labels cover every episode and were written separately from the robot operators, then audited by another team.
The partners also released FineART-VLA, a vision-language-action model based on pi0.5. It combines a higher-level component that predicts the next subtask in language with a lower-level component that turns that subtask into robot actions. The dataset is available under CC BY 4.0, while the model uses the Gemma license.
Large, richly labeled physical-action data remains a bottleneck for open robotics research. FineART gives developers a common resource for training and testing systems that need to reason over a long task rather than imitate a single motion. The release is available through Hugging Face’s LeRobot ecosystem together with code and a research paper.
