Egocentric
Egocentric demonstration datasets
First-person human manipulation for imitation learning. Structured capture, not random POV footage.
Why egocentric
Third-person cameras miss hands, contact, and the pixels a robot sees when it reaches. Egocentric capture records that view in real workplaces and homes.
Egocentric vs teleop
| Factor | Egocentric | Teleop |
|---|---|---|
| Scalability | Higher | Hardware-limited |
| Embodiment match | Transfer required | Direct |
| Natural contact | Strong | Gripper-dependent |
| Scene diversity | Strong | Cell-limited |
Full write-up: Egocentric vs teleop for robot learning
Capture setup
- Head and/or wrist POV rigs
- Optional depth and IMU
- Versioned task scripts
- Action labels and episode metadata
FAQ
Common questions
- What is an egocentric robotics dataset?
An egocentric robotics dataset captures manipulation from the demonstrator’s viewpoint, usually via head-mounted or wrist-mounted cameras. It emphasizes hand-object contact and the first-person pixels a policy may need for imitation learning.
- Egocentric or teleop?
Egocentric scales natural human demos across real scenes. Teleop maps actions directly to a robot embodiment. Many teams use both. See our comparison article for the decision frame.
- Can egocentric data include RGB-D?
Yes. We can record synchronized depth with egocentric or multi-camera setups when your model inputs require it.