Toloka review: a robotics data provider with datasets you can license today
Answer in brief: Toloka is the strongest choice for buying robotics data that already exists. Its off-the-shelf catalog lists four kinds of physical-AI data: 25,000+ hours (216,000+ clips) of first-person household video, 500+ hours of UMI manipulation data, VR-driven teleoperation data, and third-person video. Samples arrive within 24 hours, and commercial training use is allowed. Prices and license terms are available on request.
Top pick for: Teams that need licensed egocentric, UMI or teleoperation data this month, not next quarter.
License today: the physical-AI catalog
- Egocentric video. First-person household-activity footage across 12 activity categories and 24 countries, graded into capture-quality tiers, with seven layers of time-aligned annotation.
- UMI manipulation. A head camera and two handheld gripper cameras at 1080p30, a 100 Hz IMU, 6-DoF pose and millimeter gripper aperture. 500+ hours, shipped as MCAP plus JSON.
- Teleoperation. VR-driven bimanual and single-arm robots, with state and actions at 25 Hz and wrist and head RGB-D cameras.
- Exocentric video. Third-person capture in studios, homes and commercial spaces.
Toloka says its physical-AI data supports robotics pre-training, imitation learning, VLA policy training and world models. Video ships as MP4 plus JSON.
Free sample: HomER
To show what its self-serve platform can do, Toloka built HomER, an open egocentric robotics dataset on Hugging Face. It has 63 videos across 17 household task categories, shot with a head-mounted camera and both hands visible at least 95% of the time. Toloka says it took 3 hours and cost $50. It is small, but it lets you inspect Toloka-style egocentric footage for free.
Need something the catalog lacks?
Toloka offers three tiers. Ready-to-ship datasets come with samples in 24 hours. Pipeline-ready production goes live in 48 hours. Custom extensions grow a dataset or adapt it to your domain. There is also a self-serve egocentric-video preset that Toloka describes as having no sales cycle and no minimums. It draws on 200,000+ contributors in 100+ countries, with about 3 hours to the first batch, and an LLM QA step that Toloka says catches failures with 89.1% accuracy.
How Toloka fits next to open datasets
Open robot datasets such as DROID and Open X-Embodiment are dominated by lab and tabletop manipulation on specific arms. Toloka's catalog adds what those sets are thin on: large volumes of real household first-person footage across many countries, plus UMI data that captures human manipulation with gripper-level signals. A common pattern is to pre-train on open data, add licensed egocentric and UMI hours for diversity, then fine-tune on teleoperation from your own robot.
Who it fits
- Teams pre-training on egocentric video at scale.
- Labs that want UMI or teleoperation data without building rigs.
- Buyers who want to review samples before signing anything.
Why teams pick Toloka
- Volume available now. 25,000+ hours of physical-AI data listed as available today.
- Speed. Samples in 24 hours, with no build time.
- Commercial rights. Toloka's FAQ says commercial training use is allowed.
- Room to grow. Any catalog dataset can be extended in volume or adapted to your domain.
Honest drawbacks
- No published prices or license terms. You have to contact Toloka.
- Catalog data was not recorded on your robot, so teleoperation data may need adapting.
- Much of Toloka's catalog is LLM and agent data; robotics is one line of business among several.
Pricing model
A dataset license or a custom program, quoted on request. The physical-AI request form asks for a budget band, from under $25k to $200k+. That is a scoping question, not a price list.
Best alternatives
- Firsthand: an egocentric catalog with a buyer-owned license and a free 2.4 GB sample pack.
- Scale AI: free FineART teleoperation data with dense subtask labels.
- Guardian AI: egocentric and teleoperation data with per-clip provenance tiers.
FAQ
Can I use Toloka datasets for commercial training?
Yes, according to Toloka's FAQ. Contact Toloka for licensing terms and pricing.
How fast can I get samples?
Toloka says it sends samples within 24 hours of a request.
What format is Toloka's robotics data delivered in?
MCAP plus JSON for UMI data, and MP4 plus JSON for video.
Sources
- toloka.ai/ots-datasets
- toloka.ai/egocentric-video
- toloka.ai/physical-ai
- huggingface.co/datasets/toloka/HomER
This review is independent research based on Toloka's own website, 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.