Data strategy

Egocentric vs exocentric robotics data.

Camera perspective changes what a model can learn. Use this comparison to choose the right first-person, third-person, or synchronized multi-view collection for your robot task.

Inspect sample data
Decision pointEgocentricExocentric
What it sees

The demonstrator's working view and hand-object interaction.

The full workspace, body motion, and surrounding context.

Best for

Learning hand intent, contact, reach, grasp, and fine manipulation.

Learning spatial relationships, task context, body motion, and scene state.

Main limitation

Hands or objects can occlude the workspace during difficult actions.

The camera may miss the fine visual details the operator uses.

Typical placement

Headset or operator-worn stereo rig.

Tripod, overhead, side, or workstation-mounted camera.

Strongest program

First-person policy learning and hand-object interaction.

Context-rich behavior modeling and offline evaluation.

Collection guidance

Choose for the learning signal, not the camera label.

Egocentric footage is strongest when the model needs to understand what the operator sees while manipulating an object. It preserves hand intent and contact detail, but difficult actions can hide the workspace.

Exocentric footage is strongest when the model or reviewer needs the full scene: workspace layout, body motion, object location, and the relationship between actions. It can lose the close visual detail available to the operator.

When the task is high-value or hard to reproduce, synchronized views are often the most efficient choice. They let engineering teams inspect both interaction detail and full task context before scaling.

Define the model input before capture
Review occlusion and task completeness on a pilot
Keep timestamps and clip IDs aligned across views
Deliver camera specs with the annotations

Common questions

Which view should a robotics team collect first?

Start with the view closest to the model's intended input, then add the complementary view if workspace context or hand occlusion limits evaluation.

Are synchronized views better than one camera?

For many manipulation programs, synchronized egocentric and exocentric views provide both fine interaction detail and task context.

Can the same task use both camera types?

Yes. A shared hardware clock and stable clip identifiers make multi-view capture easier to align and audit.

See the difference in a real sample.

Review synchronized capture, task labels, and delivery metadata before you choose a collection plan.

Open sample explorer