GUIDE · 2026
Egocentric robot training data in 2026: what exists, what it costs you, and where the licenses bite.
Human demonstration data — first-person video with hand pose — has become a first-class training input for robot foundation models, dexterous hand policies, and world models. The sources below are what a robotics team can actually get today. The two questions that matter for each: can you use it commercially, and does it carry the signals your training recipe needs.
EgoDex (Apple)
829 hours of egocentric video with paired 3D pose for 25 joints per hand, captured on Vision Pro. 338K demonstrations across 194 tabletop tasks.
The largest public hand-pose corpus; clean synchronized pose; natural language annotations.
CC-BY-NC-ND license — research only, no commercial use, no derivatives. Tabletop tasks, not industrial work.
Ego4D / Ego-Exo4D (Meta)
Thousands of hours of in-the-wild egocentric video of daily activity; Ego-Exo4D adds synchronized exocentric views and skilled-activity focus.
Scale and diversity; the standard pretraining corpus for egocentric representations and world models.
License agreement gates use; most footage has no metric hand pose or depth; not task-specifiable.
HOT3D (Meta)
Egocentric hand-and-object 3D tracking dataset from Aria and Quest devices.
Precise hand-object interaction ground truth; good for pose and contact research.
Research license; limited hours and object set; not built for policy training volume.
Open X-Embodiment / DROID
Large collections of teleoperated robot trajectories (1M+ episodes across embodiments; DROID adds 76K episodes across hundreds of scenes).
Real robot action labels — directly consumable by VLA and imitation trainers without any bridge.
Robot data, not human: task diversity and dexterity are bounded by what teleop rigs can do; contact-rich fine manipulation is thin.
Commercial custom capture (Syngraph)
Episodes of skilled human manual work captured to a buyer's specification in industrial environments: 60fps egocentric video, LiDAR depth, 6DoF hand pose, delivered LeRobot native.
Commercial rights with per-clip consent and lineage; task and volume set by the buyer; a 7-check integrity gate with a buyer-set acceptance threshold.
No shipped catalog — the first pilot is captured against your spec. Tactile force sensing is on the roadmap, not shipping.
QUESTIONS
Can I train a commercial robot policy on EgoDex or Ego4D?
Check the license before anything else. EgoDex is CC-BY-NC-ND (non-commercial, no derivatives); Ego4D requires a license agreement. Research use is generally fine; shipping a commercial product trained on them is a legal question your counsel should answer, not a dataset card.
Human video has no robot actions. How do models use it?
Through a bridge: treat retargeted hand pose as the action stream (the recipe used by leading robot foundation models), learn latent actions from frame pairs, or align future representations without action labels. Teleop-native trainers need the retargeting step.
What format do these ship in?
Public datasets each have their own format; the robot-learning ecosystem is converging on LeRobot datasets (Parquet + MP4). Syngraph delivers LeRobot v2.1/v3.0 natively, with RLDS, HDF5, or Zarr built on spec.
How do I judge quality?
Ask for per-episode integrity checks, synchronization confidence, and provenance. Syngraph episodes pass a 7-check gate and every episode ships with a sync confidence score; you set the acceptance threshold.