FineART
FineART is a large-scale, open-source bimanual manipulation dataset co-released by Scale AI and Hugging Face LeRobot. It features 40,543 teleoperated episodes totaling 1,718 hours of high-quality robotic data, captured on Trossen Stationary AI hardware. The dataset is uniquely distinguished by its dense, human-written subtask annotations, comprising 533,913 labeled steps across 151 distinct tasks. Designed for the LeRobot ecosystem, FineART provides a comprehensive resource for training vision-language-action (VLA) models that require step-by-step reasoning and execution. The data includes multi-view camera footage, joint states, and action sequences, making it ideal for researchers and practitioners focused on contact-rich manipulation, long-horizon task planning, and cross-embodiment transfer. By providing both the raw trajectory data and structured subtask metadata, FineART enables the development of policies that can interpret goals and predict subtasks, significantly advancing the state of the art in physical AI and robotic learning.
| Year | 2026 |
|---|---|
| Episodes | 40,543 |
| Total hours | 1,718 |
| Trajectories | 40,543 |
| Frames | 185,534,500 |
| Frame rate | 30 fps |
| Embodiments | Trossen Stationary AI |
| Task categories | manipulation, bimanual, pick-and-place, long-horizon |
| License | CC-BY-4.0 |
| Access | open — commercial use permitted |
| Maintainer | Scale AI and Hugging Face LeRobot |
| Origin country | US |
FineART is a comprehensive robotic dataset developed by Scale AI's Physical AI unit and Hugging Face. It addresses the scarcity of high-quality, annotated data in robotics by providing 1,718 hours of bimanual manipulation demonstrations. The dataset is structured to support advanced VLA training, where models learn to decompose complex tasks into actionable subtasks. The collection process involved a rigorous three-stage pipeline: teleoperation for data collection, human-led temporal segmentation for subtask annotation, and a final quality validation phase. The dataset is formatted for the LeRobot library, facilitating seamless integration for training and evaluation. Use cases include training models for contact-rich skills like cable routing, precision insertion, and tool use, as well as general pick-and-place tasks. The release also includes the FineART-VLA model, which demonstrates improved performance in spatial instruction following and long-horizon task completion by leveraging the dense subtask labels provided in the dataset.
Frequently asked questions
What is the primary hardware used for FineART?
The dataset was collected using the Trossen Stationary AI, an ALOHA-style bimanual robot platform.
How is the data annotated?
Every episode includes dense, human-written subtask annotations that temporally segment the trajectory into descriptive, natural-language steps.
Is the dataset suitable for commercial use?
Yes, the dataset is released under the CC-BY-4.0 license, which permits commercial use.
How can I access the FineART dataset?
The dataset is publicly available on the Hugging Face Hub and can be loaded using the LeRobot library.
What is the relationship between FineART and FineART-VLA?
FineART is the training dataset, while FineART-VLA is a vision-language-action model trained on that data to perform step-by-step reasoning and execution.