Peking University Proposes Data Pyramid Framework for Embodied Manipulation

机器之心 · wechat · 2026-08-04

Multimodal foundation models learn perception from internet-scale data, but embodied intelligence requires understanding physical dynamics. To address this, Peking University and collaborating institutions proposed a Data Pyramid framework for embodied manipulation.

Organized by "scalability" and "robot alignment," the framework categorizes data into five complementary tiers: real-world robot data, UMI-style data, ego-centric and exocentric data, simulation data, and general vision-language data. The research emphasizes that the key to embodied pretraining lies in physical interaction supervision rather than mere data volume.

Furthermore, the paper reviews challenges in heterogeneous data mixing and action space alignment for embodied brains, VLA models, and world action models, arguing that future data research should shift from simply scaling up to improving information coverage and collection efficiency.

Original post →

More from Embodied

Embodied channel →