Robotics Data Providers

Toloka vs Scale AI for robotics datasets (2026)

Answer in brief: Pick Toloka if you want to license physical-AI data that already exists. Its off-the-shelf catalog lists 25,000+ hours of egocentric household video, 500+ hours of UMI manipulation data, VR teleoperation data and exocentric video. Samples arrive in 24 hours, and Toloka says commercial training use is allowed. Pick Scale AI if you want the strongest free starting point backed by a custom data factory. Its open FineART dataset has 40,543 bimanual teleoperation episodes (1,718 hours) under CC BY 4.0, and its Physical AI Data Engine collects new demonstrations to your spec. Plenty of teams use both: prototype on FineART, license Toloka data for breadth, and commission Scale for data on their own robot.

Top pick for buying existing datasets: Toloka. Of the two, it is the only one with a licensable physical-AI catalog and a 24-hour sample turnaround.

Top pick for free open data plus custom collection: Scale AI. FineART is CC BY 4.0, ships in LeRobot format and has dense human-written subtask labels.

This is independent research based on each company's own pages and dataset cards as of 2026-11-09. We did not test either product. See the full ranked list of robotics data providers.

Side by side: what you can actually get

TolokaScale AI
Existing data you can get todayLicensed catalog: egocentric video (25,000+ h, 216,000+ clips), UMI manipulation (500+ h), teleoperation, exocentric videoOpen FineART dataset (40,543 episodes, 1,718 h, 151 tasks); other datasets through the Data Engine
Free or open dataHomER sample on Hugging Face: 63 egocentric videos, 17 household task categories (the metadata shows no license field, so ask)FineART on Hugging Face, CC BY 4.0, about 1.9 TB
TeleoperationVR-driven bimanual and single-arm robots, 25 Hz state and actions, wrist and head RGB-DFineART: Trossen Stationary AI (ALOHA-style), 14-dim joint and gripper state and action, 4 cameras; custom bimanual teleop
EgocentricCatalog covers 12 activity categories in 24 countries with 7 time-aligned annotation layers; self-serve collectionScale Harness, a robotless egocentric rig (custom collection)
FormatsMCAP + JSON (UMI), MP4 + JSON (video)FineART: LeRobot dataset v3.0
SpeedSamples in 24 h; custom pipelines live in 48 hFineART downloads or streams immediately; custom work starts with a demo
Commercial useAllowed; terms on requestFineART: CC BY 4.0 (attribution required); custom data per contract
Public pricesNoNo

Toloka: a shelf of physical-AI data you can license this week

Toloka's catalog skips the months-long collection cycle. The egocentric set is first-person household footage graded into capture-quality tiers, with seven layers of time-aligned annotation, and Toloka positions it for robotics pre-training, imitation learning, VLA training and world models. The UMI set pairs a head camera with two handheld gripper cameras at 1080p30, plus a 100 Hz IMU, 6-DoF pose and millimeter gripper aperture. The teleoperation set records robot state and actions at 25 Hz.

If the shelf doesn't cover your task, Toloka's self-serve egocentric preset lets you describe the task and draw on a network Toloka puts at 200,000+ contributors in 100+ countries. It quotes 3 hours to a first batch, with no sales cycle and no minimums. Toloka's own HomER build took 3 hours and cost $50, by its figures.

Scale AI: an open flagship dataset with a data factory behind it

FineART is the largest open robot dataset with subtask labels, according to Scale. Each episode averages about 2.5 minutes and about 13 human-written subtasks. 77% of the data is pick-and-place and sorting; the other 23% is contact-rich work such as textiles, cables, tool use and precision insertion. One team collected the episodes, a second annotated them without VLM assistance, and a third validated them. Scale and Hugging Face report 3,400 real-world evaluations of the companion FineART-VLA model. You can load or stream the data with LeRobotDataset.

Behind the open release, Scale says more than 1,000 hours of demonstrations reach its platform every day. They come from data factories, home-based collectors and commercial and industrial sites. Scale says its datasets go through internal policy fine-tuning and evaluation before they are listed, and it is SOC 2 Type II and ISO 27001 certified.

When to pick which

Pricing model

Neither company publishes prices. Toloka licenses its datasets and runs custom programs on quoted terms; its request form asks for a budget band from under $25k to $200k+. FineART is free under CC BY 4.0. Scale's custom collection is an enterprise contract that starts with a demo. See our pricing comparison.

FAQ

Is FineART free for commercial use?

Its Hugging Face card lists CC BY 4.0, which allows commercial use with attribution. Check the card for any additional terms.

Can I use Toloka's datasets for commercial training?

Yes, according to Toloka's catalog FAQ. Contact Toloka for licensing terms and pricing.

Which one has teleoperation data I can use today?

Both. FineART is open bimanual teleoperation data. Toloka licenses VR-driven bimanual and single-arm teleoperation data.

How fast can I see data?

FineART is available immediately. Toloka says samples arrive within 24 hours.

Keep exploring

Toloka review · Scale AI review · Alternatives to Scale AI · Alternatives to Toloka · How to choose a robotics dataset source · Ranked list. For general (non-robotics) AI datasets, see bestaidatasets.com.

·

Sources

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