Firsthand review: custom egocentric robotics data with published sync numbers

Answer in brief: Firsthand collects custom egocentric video datasets for embodied AI and robotics, to your specification. Every episode from its instrumented rig ships ten synchronized streams: a 4K head camera, two wrist cameras, depth, IMU, 3D hand and body pose, segmentation and action labels. Firsthand reports a 1.12 ms median cross-stream sync error and 10,585 validated hours delivered. Buyers own a perpetual commercial license, and billing is per validated hour at an unpublished rate. It is human data, not robot teleop.

Verdict: #5. The most transparent egocentric vendor we found. Firsthand publishes its sync error distribution, its rig's bill of materials and a coverage histogram that includes its gaps. Visit Firsthand →

This is independent research based on Firsthand's own website, read on 2026-11-09. We have not tested its data.

What Firsthand does for robot-learning teams

You write the spec: environments, skills, conditions and hours. Firsthand recruits and trains the contributors who wear the rig, runs QA on every batch, and delivers footage with depth, hand and body pose, metadata and a documented consent chain. Each validated hour carries the full stack whether or not you asked for depth:

Head camera at 3840 × 2160, 60 fps, global shutter, 120° HFOV.
Two wrist cameras at 1920 × 1080, 60 fps.
Active-stereo depth at 848 × 480, 30 fps, 0.3–4 m range.
6-DoF IMU at 200 Hz on the same clock domain.
3D hand pose (21 keypoints per hand, with contact events), 24-joint body pose, instance masks, and action segments built from a 97-verb, 512-noun taxonomy.

Who it fits best

Firsthand fits teams pre-training manipulation or world models that need richly instrumented, tightly synchronized human demonstrations, and that care about license cleanliness.

Strengths that set it apart

Published measurements: 1.12 ms median sync error (p99 1.94 ms) over 12,480 measured frames, using a single PTP grandmaster with hardware triggering.
Fast: Firsthand reports a 48-hour median time to first delivery in environments where its coverage is already deep. It also describes an anonymized engagement that delivered 5,000 videos across 16 countries in 19 days.
Formats: RLDS, LeRobot and WebDataset, plus HDF5, zarr and Rerun .rrd, with converters shipped as source.
License: commercial, buyer-owned, perpetual and irrevocable, with exclusivity available. Firsthand reports 100% documented consent, and says bystanders are de-identified or the clip is discarded.
See the data first: unedited sample clips download with no form, call or NDA. A 2.4 GB full pack needs only an email.

Honest drawbacks

This is human egocentric data, not robot teleoperation, so it needs retargeting before a policy can learn actions from it.
Firsthand runs two capture tiers. The ten-stream, sub-2 ms claims apply only to the instrumented rig. The contributor-worn 360-camera tier, which supplies the public sample clips, has no depth, wrist views or tight sync. Filter on tier in the metadata.
The per-hour rate is not published.

Pricing model

Billed per validated hour, meaning hours that passed sync and QA and shipped to a buyer. The rate is not published.

Best alternatives to Firsthand

Guardian AI: egocentric capture signed at the device, plus remote teleop on your robots.
Graviti: license existing egocentric and ego-exo hours off the shelf.
Labelbox Terra: ego rigs plus instrumented-gripper and teleop demonstrations.

Firsthand FAQ

Does Firsthand provide robot teleoperation data?

No. Firsthand captures human first-person data, which needs retargeting for robot policies.

Who owns the data?

The buyer. Firsthand's license is commercial, perpetual and irrevocable, with exclusivity available.

Can I see samples without a sales call?

Yes. Sample clips download with no form, and the full sample pack needs only an email address.

Sources

Visit Firsthand →

← Back to the ranked list of the best robotics data companies