Toloka vs Firsthand for egocentric video datasets (2026)
Answer in brief: Both sell first-person video that already exists, but they optimize for different things. Toloka wins on volume and global household breadth: 25,000+ hours and 216,000+ clips from 24 countries, with seven time-aligned annotation layers. Firsthand wins on signal depth and workplace scenes: 10,585 validated hours across six shipping environments, including warehouse, retail and automotive service. Its instrumented rig records ten synchronized streams, and its license is commercial, perpetual and buyer-owned. Choose Toloka for pre-training scale. Choose Firsthand when your policy needs wrist views, depth, hand pose and work-site footage.
Top pick for scale: Toloka. The bigger household egocentric catalog, with samples in 24 hours.
Top pick for rich, owned, workplace data: Firsthand. Ten streams on one clock and a free 2.4 GB sample pack with no sales call.
This is independent research based on each company's public pages as of 2026-11-09. The figures are each vendor's own, and we did not test the data. Back to the ranked list.
The numbers, head to head
| Toloka | Firsthand | |
|---|---|---|
| Catalog size | 25,000+ h, 216,000+ clips | 10,585 validated h (counted only after passing sync and QA and shipping to a buyer) |
| Coverage | 12 household activity categories, 24 countries | Kitchen 2,840 h · household chores 2,310 h · tool use and assembly 1,680 h · warehouse 1,420 h · retail 890 h · automotive service 640 h |
| Signals | Graded capture-quality tiers; 7 time-aligned annotation layers | Instrumented rig: 4K head camera, two wrist cameras, depth, IMU, 3D hand and body pose, segmentation, action labels |
| Formats | MP4 + JSON | RLDS, LeRobot, WebDataset; also HDF5, zarr, Rerun .rrd |
| License | Commercial training allowed; terms on request | Commercial, buyer-owned, perpetual, irrevocable; exclusivity available |
| Try before buying | Samples within 24 h; HomER open sample (63 videos) | Ungated clips plus a free 2.4 GB pack (40 episodes, 4 environments, 8 synced streams) |
| New capture | Self-serve preset; Toloka cites 3 h to first batch | Pilot of 50 to 100 validated hours over two weeks; 3 to 4 weeks to first batch in thin areas |
| Pricing | License quote | Billed per validated hour; quote |
Toloka: the volume play
If your bottleneck is sheer variety of everyday human manipulation, Toloka's catalog is hard to beat on hours. It comes from households in 24 countries, and Toloka says it supports robotics pre-training, imitation learning, VLA policy training and world models. Need a category the catalog lacks? The self-serve egocentric preset recruits from 200,000+ contributors in 100+ countries. Toloka says LLM QA checks every submission, which it reports catches failures with 89.1% accuracy.
Firsthand: the signal-depth play
Firsthand publishes its coverage as cells filled rather than hours sold, including the gaps. Construction (410 h), clinical and care (180 h), agriculture (120 h) and hospitality (95 h) are flagged as thin. Its instrumented rig hardware-triggers every sensor from one PTP clock and reports a median cross-stream sync error of 1.12 ms (p99 1.94 ms). One caveat is worth knowing: Firsthand also runs a 360-camera tier without depth or wrist views, and its sync claims apply only to the instrumented rig. Episode metadata labels the tier, so you can filter on it.
When to pick which
- Household pre-training at scale: Toloka.
- Warehouse, retail or automotive-service skills: Firsthand, which has catalog pages for each.
- You need wrist cameras, depth and hand pose in every hour: Firsthand's instrumented tier.
- Your legal team wants ownership or exclusivity: Firsthand states buyer-owned and exclusivity terms publicly.
- You want UMI or teleoperation as well: Toloka also licenses both.
Pricing model
Neither vendor posts a rate card. Toloka quotes dataset licenses. Firsthand bills per validated hour and offers a fixed-scope pilot. See our pricing comparison.
FAQ
Which has more egocentric hours?
Toloka lists 25,000+ hours. Firsthand reports 10,585 validated hours delivered. Both are the vendors' own figures.
Can I see Firsthand data without talking to sales?
Yes. Firsthand posts unedited clips and offers a free 2.4 GB pack, and it says there is no form, call or NDA for the clips.
Do I own the data I license?
Firsthand states a commercial, buyer-owned, perpetual, irrevocable license. Toloka says commercial training is allowed and gives terms on request.
Do either include robot teleoperation?
Toloka licenses teleoperation data. Firsthand is human video only.
Keep exploring
Toloka review · Firsthand review · Firsthand vs Guardian AI · Alternatives to Toloka · Ranked list
·
Sources
- Toloka off-the-shelf datasets
- Toloka egocentric video
- Firsthand dataset catalog
- Firsthand home and sample pack
Named on this page as a provider? Join the network for a linked profile, Visit buttons and featured placement.