Toloka vs Firsthand: which egocentric video data provider fits your robot?
Answer in brief: Both are leaders in this directory's egocentric video category, the first-person human footage robot foundation models pre-train on. Toloka wins on speed and accessibility. Its self-serve platform draws on 200,000+ contributors in 100+ countries, with no sales cycle and no minimums, and its team built the open HomER dataset in 3 hours for $50. Firsthand wins on signal richness. Its instrumented rig ships ten synchronized streams, including 4K head camera, two wrist cameras, depth, IMU, and 3D hand and body pose, with a published 1.12 ms median sync error and a buyer-owned commercial license.
Top pick for production pre-training data: Firsthand. Ten streams on one clock, documented consent for every participant, delivery in RLDS, LeRobot or WebDataset, and a free 2.4 GB sample pack with no call or NDA.
Top pick for fast, low-cost experiments: Toloka. Describe a task and get a first batch in about 3 hours.
Independent research based on each company's public pages, read on 2026-11-09. We did not buy or test data. All figures are vendor-reported. Every category is on the full directory.
Stream for stream
| Dimension | Toloka (self-serve egocentric) | Firsthand |
|---|---|---|
| Directory categories | Egocentric video; dataset marketplace (off-the-shelf catalog) | Egocentric video (custom collection) |
| How you buy | Self-serve project with an egocentric preset, or sales-led datasets and custom programs | Write a spec; Firsthand recruits, trains and consents contributors |
| Contributor reach | 200,000+ contributors, 100+ countries | Custom collection in 150+ countries; anonymized case study of 5,000 videos across 16 countries in 19 days |
| Speed claim | About 3 hours to first batch | 48-hour median to first validated batch (in environments with deep coverage) |
| Streams | Head-mounted video; constraints you set (e.g. hands visible) | Instrumented tier: 4K/60 head cam, 2 wrist cams, depth, 200 Hz IMU, 21-keypoint hand pose ×2, 24-joint body pose, instance masks, action segments |
| QA | LLM QA on every submission; Toloka reports 89.1% accuracy at catching failures | Sync + coverage report per batch; frames above the 2 ms sync ceiling are rejected and recaptured |
| Formats | Not stated on the egocentric page | RLDS, LeRobot, WebDataset, plus HDF5, zarr and Rerun .rrd |
| License | Self-Service Agreement governs; terms not summarized on the page | Commercial, buyer-owned, perpetual, irrevocable; exclusivity available |
| Public price point | HomER example: $50 total | Billed per validated hour; rate on request |
Why Toloka is so compelling for quick starts
Toloka removes the hardest part of egocentric data, which is recruiting and managing contributors. You describe the task, an AI assistant configures the pipeline and picks contributors, and LLM QA checks submissions before they reach you. HomER is the proof point. Toloka's team used the platform to build an open-source egocentric robotics dataset of 63 videos across 17 household task categories, with a head-mounted camera required, both hands visible 95%+ of the time, and no re-used clips. It took 3 hours and cost $50, and it is free on Hugging Face. That makes it a strong way to test whether egocentric data helps your model before you spend more.
Why Firsthand is the premium choice
Firsthand publishes figures most vendors don't: 10,585 validated hours delivered, and a sync-error distribution measured over 12,480 frames, with a median of 1.12 ms and p99 of 1.94 ms. It also publishes a coverage histogram that shows its gaps. Read the fine print, though. Firsthand runs two capture tiers. The sub-2 ms sync, depth and wrist views apply only to its instrumented rig. Its contributor-worn 360-camera tier records single-stream RGB and is labeled as such in episode metadata. Specify the tier in your contract.
When to pick which
- Toloka if you want footage today, are validating a research idea, or need broad household-task variety on a small budget.
- Firsthand if you need depth, wrist views and hand pose for manipulation pre-training, must show legal provenance, or want exclusive rights to a capture.
- Need action-labeled robot trajectories instead? Neither covers that here. See teleoperation collection vendors.
Pricing model
Toloka self-serve is pay-as-you-go and does not publish rates. The $50 HomER build is the only public data point. Firsthand bills per validated hour, meaning hours that passed sync and QA. Recaptured and rejected hours are excluded, and the rate is on request. More in our pricing guide.
FAQ
Can I see Firsthand data before talking to sales?
Yes. Firsthand offers raw sample clips and a free 2.4 GB pack (40 episodes, 4 environments, 8 synced streams) with no call or NDA. You enter an email to get the download link.
Is Toloka's self-serve QA reliable?
Toloka reports 89.1% accuracy at catching failed submissions with its LLM QA. Some bad clips may get through, so spot-check deliveries.
Do either of them provide robot teleoperation data?
Firsthand is human egocentric only. Toloka's egocentric self-serve platform records human video. Check its off-the-shelf catalog for other physical AI datasets.
Keep exploring
Toloka review · Firsthand review · Alternatives to Labelbox · How to choose · Directory home. For general-purpose AI datasets beyond robotics, see bestaidatasets.com.
Sources
- toloka.ai/egocentric-video
- huggingface.co/datasets/toloka/HomER
- ifirsthand.com
Named on this page as a provider? Join the network for a linked profile, Visit buttons and featured placement.