Scale AI and Hugging Face release FineART robotics data

Robotics training data is getting more detail about the steps inside a task.

What happened

Scale AI and Hugging Face released FineART and the FineART-VLA model. The dataset contains 40,543 recorded episodes, covering 1,718 hours, with 533,913 labelled subtasks. The release is integrated with the LeRobot ecosystem.

Why it matters

Intermediate labels can help researchers study where multi-step execution breaks down. They also create training material that connects a high-level instruction with the actions needed to complete it.

The bigger picture

Open robotics data can reduce barriers to experimentation. Its practical value still depends on task diversity, annotation quality and how well learning transfers to unfamiliar environments.