Data Pyramid for Embodied Manipulation

Researchers propose a five-part taxonomy for embodied manipulation data: real-robot, UMI-style, egocentric/exocentric, simulation, and general vision-language data.

They compare these sources across scale, robot alignment, quality, diversity, reusability, and physical fidelity, showing clear tradeoffs for training robot models.

The paper reviews how recent embodied foundation models mix these data sources during pretraining to balance broad coverage with action realism.

It also highlights open gaps in tactile data, failure-and-recovery trajectories, scalable collection pipelines, cross-embodiment action alignment, and better use of egocentric data for dexterous skills.