Guide // Robot learning data
Egocentric video vs teleoperation data
Two common ways to gather training data for robot manipulation and humanoid imitation learning — and how teams often combine them.
At a glance
| Egocentric video | Teleoperation | |
|---|---|---|
| What is captured | People performing real tasks, filmed from their own point of view | A robot controlled by a human operator, with robot joint and action data |
| Scale | Easier to scale across many people, places, and tasks | Limited by robot hardware and operator time |
| Environment variety | Real homes, kitchens, shops, and workplaces | Usually labs or set-up workspaces |
| Direct robot actions | No — actions must be learned or inferred from video | Yes — recorded in the robot's own action space |
| Typical use | Pre-training, task understanding, imitation from human video | Fine-tuning a policy on a specific robot |
Why many teams use both
Large amounts of varied first-person human video help a model learn how tasks unfold across many real settings. Smaller sets of teleoperation demonstrations then teach the exact movements for a particular robot. The right mix depends on your model, robot, and target tasks.
Collecting multi-angle video for humanoid imitation learning
- Keep the hands and the task area in frame for the whole clip.
- Agree on camera positions and how angles are time-synced before recording starts.
- Define a task list so every clip has a clear label and start and end point.
- Record across varied homes and workplaces, not one set-up room.
- Review every clip for privacy and quality before delivery.
Veda Solutions runs egocentric data collection in India, with annotation and quality control in-house. See a sample dataset or our delivery formats.
Have a Data Collection or Annotation Project?
Whether you need real-world data collected, an existing dataset annotated, or an end-to-end data pipeline, Veda Solutions can build and scale the workflow around your requirements.