A LeRobot dataset is episodic robot learning data built for the Hugging Face LeRobot library. Each episode typically includes camera observations, actions, and metadata so teams can train imitation learning policies with a shared loading path.
What sits inside an episode
Expect synchronized observations (often RGB, sometimes depth), action vectors or teleop signals, and task metadata. Exact on-disk layout follows LeRobot’s current conventions and your chosen backend. The point is consistency: load, visualize, train, evaluate without a one-off parser.
When open datasets are enough
Public corpora help research baselines and early experiments. They are weaker when your objects, fixtures, lighting, or embodiment differ from the release. That gap is where custom demonstration collection matters.
When to commission custom LeRobot data
Commission custom data when you need proprietary skills, exclusive rights, RGB-D sync to your sensors, or packaging pinned to your LeRobot version. See our LeRobot data collection service for scope, deliverables, and packaging.