Firsthand vs Guardian AI for egocentric robotics data (2026)
Answer in brief: Choose Firsthand if you want hours that already exist. It publishes a catalog of 10,585 validated hours across six environments and a free sample pack, and it sells under a buyer-owned commercial license. Choose Guardian AI if you need proof that every clip is real and consented, or if you also want robot teleoperation. Guardian captures new data across Asia and the US, stamps each clip with a P0 to P3 provenance tier, and has its operators drive your robots remotely. Firsthand is the faster path to existing data. Guardian is the stronger path to audited new data.
Top pick for data you can download now: Firsthand. An existing catalog, ten synchronized streams and a free 2.4 GB pack.
Top pick for verified provenance plus teleop: Guardian AI. Clips signed at capture, screened for synthetic or re-filmed footage, and delivered with a manifest.
This is independent research based on each company's own site as of 2026-11-09, with no hands-on testing. Back to the ranked list.
Head to head
| Firsthand | Guardian AI | |
|---|---|---|
| Existing catalog | Yes: 10,585 validated h across 6 shipping environments | No public catalog; captures to your scope |
| Robot teleoperation | No | Yes; remote operators, and your robots stay at your site |
| Provenance | 100% documented consent (its figure); sample consent form published | Tiers P0 (declared) to P3 (authenticated physical capture), shown per clip |
| Signals and labels | Head, two wrist cameras, depth, IMU, hand and body pose, segmentation, action segments | 3D hand pose, 6-DoF camera trajectory, hand-object contact, task segmentation and narration, success, failure and recovery labels |
| Formats | RLDS, LeRobot, WebDataset, HDF5, zarr, .rrd | LeRobot, RLDS, WebDataset, HDF5, MCAP |
| Privacy | Face and license-plate blurring services | Faces, screens and bystanders anonymized before delivery |
| How you start | Free sample pack; pilot of 50 to 100 validated hours in two weeks | 30-minute call; paid pilot of at least 20 finished hours per week with written success criteria |
| Pricing | Per validated hour | Per finished hour |
Firsthand: buy what is already on the shelf
Firsthand's catalog is unusually transparent. It lists hours, episodes, participants and sites for each environment (kitchen has 2,840 h and 18,420 episodes), plus a coverage matrix that shows where it is thin, such as low light, multi-person scenes and failure cases. It reports a 48-hour median time to first delivery where coverage is deep, and 3 to 4 weeks in thin areas. Keep in mind that its multi-stream sync figures apply to the instrumented rig, not to its 360-camera tier.
Guardian AI: capture you can audit
Guardian's pitch is trust. At P2, a clip is recorded on an enrolled device and signed at capture, and every later edit is logged. At P3, motion sensors and video must agree physically, and the clip must pass synthetic and re-filming screens. Each weekly batch ships with a QA report covering hours accepted and rejected, per-layer metrics against agreed thresholds, and the provenance tier of every clip. Guardian says it runs QA within 24 hours of each session, and it lists locations in San Francisco, Korea, Taiwan and the Philippines.
When to pick which
- Kitchen, household, warehouse or retail skills needed now: Firsthand.
- Legal requires contributor-level consent evidence: Guardian at P1 to P3, or Firsthand's documented consent chain. Ask both for samples.
- You need demonstrations on your own robot: Guardian.
- You want exclusivity: Firsthand advertises exclusivity, and its pilots in thin coverage cells can buy exclusive capture.
- Recovery and failure episodes: Guardian labels deliberate recovery episodes; Firsthand says failure cases are its thinnest column.
Pricing model
Both quote per hour of finished data after scoping. Guardian requires a paid pilot of at least 20 finished hours per week. Firsthand's pilot is fixed-scope at 50 to 100 validated hours. Neither publishes a dollar rate. See our pricing comparison.
FAQ
Does Guardian AI sell existing datasets?
Its site describes scoped capture programs, not a catalog. The factory clips on its site are labeled as samples from the open Build AI Egocentric-10K dataset.
Which formats overlap?
Both deliver LeRobot, RLDS, WebDataset and HDF5. Guardian adds MCAP, and Firsthand adds zarr and Rerun .rrd.
Can Firsthand do teleoperation?
No. Firsthand collects human first-person video only.
Keep exploring
Firsthand review · Guardian AI review · Toloka vs Firsthand · How to choose · Ranked list. For a ranking of custom-collection vendors, see roboticsdatacompanies.com.
·
Sources
- Firsthand home
- Firsthand dataset catalog
- Guardian AI home
Named on this page as a provider? Join the network for a linked profile, Visit buttons and featured placement.