EmbodiedData

EMBODIEDDATA DOCUMENTATION

Dataset contribution guide

Prepare a dataset that other researchers can understand, validate, and reuse.

Before you upload

Confirm that you own the data or have distribution permission, remove secrets and confidential material, and obtain consent from identifiable people. Choose a licence compatible with the source material.

1. Organize episodes

Use stable episode identifiers. Keep observations, actions, timestamps, instructions, outcomes, and calibration data machine-readable. Separate raw and derived assets.

2. Describe the embodiment

Record robot model, degrees of freedom, end effector, control interface, observation frequency, action frequency, coordinate frames, and sensor calibration.

3. Document outcomes

Define success before labeling. Include failed demonstrations when they are useful and explain failure categories.

4. Validate honestly

Resolve corrupted files, missing timestamps, duplicate names, count mismatches, and missing metadata. Automated checks are structural aids, not certification.

Human demonstrations

Clearly mark datasets without robot actions. Human video can be valuable for representation learning, affordances, and task understanding, but is generally insufficient for direct action-policy training.

Explore reviewed datasets →