Egocentric video
Hands, tools and context from the worker’s view.
Scope: Viewpoint, visibility and scene variation.
For robotics labs
Robot training data grounded in real work.
First-person recordings and annotations for robotics labs, world-model teams and research groups. Scope the task, capture method and data your model needs.
Choose your capture method.
Robotics data collection starts with the signal your model needs.
Human demonstrations
A task shown from setup to outcome.
Scope: Instructions, action steps and recovery.
Teleoperation data
Robot observations and commands during human control.
Scope: Embodiment, actions and synchronized signals.
Define each episode.
Agree the specification before egocentric data collection begins.
- Task & outcome
- Instruction, starting state and completion criteria.
- Episode boundaries
- Start, end, resets and partial attempts.
- Video & timing
- Viewpoint, frame rate, resolution and synchronization.
- Coverage & quality
- People, settings, variations and acceptance checks.
Choose the annotations.
Agree the fields and label definitions your project needs.
- Episode & source IDs
- Trace each segment to its recording.
- Actions & objects
- Timestamp steps and visible state changes.
- Hands & outcomes
- Scope tracks, completion, failure and recovery.
- Quality & versions
- Record uncertainty and label revisions.
Specify your training format.
LeRobot, HDF5 and RLDS are requirements to scope, not verified integrations. Conversion cannot create unrecorded robot actions.
LeRobot
Dataset version, features, camera keys and action representation.
HDF5
Groups, shapes, types, units and compression.
RLDS
Observations, actions, step flags and episode metadata.
Questions, answered.
What is robot training data?
Robot training data contains observations, demonstrations, instructions and, where recorded, actions for training or evaluating robot models. The required signals depend on the task and robot.
Does egocentric video include robot actions?
No. Egocentric video records a person’s view. Robot actions need separately measured signals or a justified mapping.
Can I request LeRobot, HDF5 or RLDS?
Yes. Include the format and loader version in your inquiry. Wake scopes available fields and compatibility before agreeing delivery.
What should your robot learn?
Share the task and data requirements. Start here or use our contact form.
Format references: Hugging Face: LeRobotDataset v3.0 · Google Research: RLDS data model