What it is
Robot data differs from web text or images because it must encode action and physical interaction.
Why it matters
Data scarcity and heterogeneity are central constraints for general robot learning.
How not to overread it
More data is not automatically better if embodiments, tasks, sensors, or labels are incompatible.
Related edges
Vision-language-action models
Robot learning
Web pretraining does not replace robot action data.
Cross-embodiment data
Dataset composition
Cross-embodiment is not automatic cross-hardware reliability.
LeRobot
Open robot learning workflows
Tooling is not a benchmark result.
Synthetic data
Dataset expansion
Synthetic data must not be counted as real-world proof.