ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine

Researchers introduced ACE, a human-centric ambient capture system that synchronizes egocentric and exocentric video with body, hand, object-state, audio, and touch signals in real homes.

Using ACE, they built ACE-Data-0: 150 hours of data, 17 million frames, 50 participants, 200 task categories, and 75,000 interaction episodes across two environments.

The dataset provides aligned demonstrations of perception, motion, manipulation, and interaction over time, filling a major gap in embodied AI training data.

A hierarchical benchmark shows current methods still struggle with contact, occlusion, egomotion, and long-horizon tasks.

The resource could improve imitation learning, world models, and vision-language-action systems with richer supervision than existing fragmented datasets.