New survey maps embodied manipulation data into a five-layer pyramid
liuziwei7 · x · 2026-07-28
Data Pyramid for embodied manipulation
A new survey proposes a five-layer data pyramid for embodied foundation models, arguing that embodied AI cannot rely on internet-scale image-text data alone because it needs data that couple perception, physical state, and actions.
The pyramid spans:
- Real-robot data: best alignment with physical reality, but hardest to scale
- UMI-style data: structured manipulation demonstrations
- Egocentric / exocentric data: broader but less physically faithful
- Simulation: scalable and reusable, with a sim2real gap
- General vision-language data: cheapest and most abundant, but weakest robot alignment
The paper also analyzes recent embodied foundation models through the lens of their data mixtures, relating data composition to perception, reasoning, planning, action generation, and world prediction. It closes by listing open problems such as tactile datasets, better data collection, and how to balance scalability against robot alignment.
Related event: Survey Proposes Five-Tier Data Pyramid for Embodied Control(2 posts)→
More from Embodied
- Origin Autonomy emerges with $20M to automate interior construction, robots running 3x human speed — lukas_m_ziegler · 2026-10-17
- One-piece 4-meter PETG bar counter shows how far large-format 3D printing has come — lukas_m_ziegler · 2026-10-17
- PayBox open-source SDK gives robots a wallet — a hexapod paid real USDC on-chain — R4nsack · 2026-10-17
- JR West to deploy giant humanoid robots this fall, cutting high-altitude rail crews by 30% — CyberRobooo · 2026-10-17
- Hobbyist models humanoid robot, hits battery space and servo torque limits — _Stocko_ · 2026-10-17
- Android robots learn to switch emotions in real time, crossing the uncanny valley — TinfoilTricorn · 2026-10-17