Scale AI review: a robotics data provider with a free dataset and a data engine

Answer in brief: Scale AI is the best place to start if you want a large, densely labeled robot dataset today and a supplier who can scale with you later. Scale AI and Hugging Face LeRobot co-released FineART under CC-BY-4.0. It contains 40,543 teleoperated bimanual episodes (1,718 hours, 151 tasks), and every episode is annotated with human-written subtasks, for 533,913 subtask labels in total. Beyond FineART, Scale's Physical AI Data Engine sells custom demonstration data under enterprise contract. There is no public price.

Top pick for: VLA and hierarchical-policy teams that want free, step-by-step language-labeled demonstrations now, plus a supplier that can grow with them.

Download today: FineART

FineART is on Hugging Face as lerobot/FineART, in LeRobot dataset v3.0 format, 1.86 TB in total. Every episode was recorded on a Trossen Stationary AI (ALOHA-style) robot with ALOHA2 grippers. Each frame has a 14-dimensional joint-and-gripper state and action, plus four cameras (bird's-eye, worm's-eye and both wrists) at 360x640 RGB, 30 fps.

Most robot datasets give one instruction per episode. FineART annotates every episode in detail. Episodes average about 2.5 minutes and about 13 subtasks. The dataset card says the annotators wrote labels without VLM assistance, worked separately from the teleoperators, and were audited by a third team. About 77% of hours are everyday pick-and-place and sorting. The rest is a long tail of contact-rich skills: deformable objects, tool use, cable routing, compliant alignment and precision insertion.

You don't have to download it all. LeRobot can load selected episodes, or stream the dataset with StreamingLeRobotDataset. Scale and Hugging Face also released FineART-VLA, an open model trained on these annotations.

When open data runs out: the Physical AI Data Engine

Scale collects to order three ways: teleoperated bimanual manipulators; Scale Harness, a first-person rig with no robot for distributed dexterous collection; and your own hardware, managed and operated by Scale. Collection runs in centralized data factories, in collectors' homes, and at commercial and industrial sites. Scale says more than 1,000 hours of demonstration data reach its platform every day, that collection cells are developed and validated in its San Francisco R&D lab before deployment, and that datasets go through internal policy fine-tuning and evaluation before they are listed. For a deeper comparison of custom-collection vendors, see our sister guide on robotics data companies.

What partners say on Scale's page

Scale's physical-AI page quotes Brad Porter: “At Cobot, our mission is to deploy physical AI at scale through our robot, Proxie… We're excited by the work Scale has already done in building this data set and look forward to partnering to push the data frontier in physical AI.” It also quotes Pete Florence calling Scale “a trusted partner in making abundant, high-quality, custom data.”

Who it fits

• Foundation-model and VLA teams that want free, step-by-step language labels to train on immediately.
• Labs that plan to move from open data to custom, multi-embodiment collection with the same supplier.
• Enterprises that need certified handling. Scale says it is SOC 2 Type II and ISO 27001 certified, supports GDPR, CCPA and customer audit requirements, and offers optional onshore data processing.

Why teams pick Scale

• Dense labels for free. 533,913 subtask labels is unusually deep annotation for an open dataset.
• Published results. The FineART authors report that mid-training on FineART raised success on a spatial-instruction task from 32% to 100%, and on a long-horizon task with human subtask guidance from 16% to 76%. On a new robot (YAM), it matched baseline performance with about 10x less fine-tuning data. These are the authors' own figures; see the paper.
• Named partners. Scale names Generalist and Physical Intelligence as partners on its data platform.
• In-house roboticists help write data specifications for custom programs.

Honest drawbacks

• FineART covers one robot type at 360x640 resolution. If you need multiple embodiments or higher resolution, you will need more data.
• The task mix is uneven. Sample by task_index if you want the contact-rich long tail represented.
• The data engine has no public catalog, price list or standard license, and its volume figures are Scale's own.

Pricing model

FineART is free under CC-BY-4.0, which requires attribution. Data-engine work is an enterprise contract; book a demo to get a quote.

Best alternatives

• Toloka: license existing teleoperation, UMI and egocentric data, with samples in 24 hours.
• NVIDIA Isaac Sim: free simulation plus NVIDIA's open Physical AI datasets.
• Encord: collection in its own facilities plus a multi-sensor annotation platform.

FAQ

Is FineART free for commercial use?

The Hugging Face card lists the license as CC-BY-4.0, which permits commercial use with attribution. Read the license yourself before you ship a product.

What robot was FineART recorded on?

A Trossen Stationary AI bimanual robot with ALOHA2 grippers, with 14-dimensional joint and gripper state and action.

Does Scale sell a catalog of ready-made datasets?

Scale's page says datasets are validated before they are listed on its platform, but it publishes no catalog or prices. Book a demo to see what is available.

Sources

• scale.com/physical-ai
• huggingface.co/datasets/lerobot/FineART
• arxiv.org/abs/2609.36416

This review is independent research based on Scale AI's own website and the official dataset card, checked on 2026-11-09. We did not test the product.

← Back to the ranked list of robotics data providers

Named on this page as a provider? Join the network for a linked profile, Visit buttons and featured placement.