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

TL;DR - ACE-Data-0 is a synchronized multisensory dataset capturing human activity at table and room scales. It addresses fragmented embodied-AI supervision for imitation learning, world models, and vision-language-action systems.

• Contains 150 hours, 17 million video frames, and 75,000 episodes across 200 tasks performed by 50 participants.
• Aligns egocentric and exocentric video with body and hand motion, object geometry and trajectories, audio, and tactile signals.
• Covers fine-grained manipulation, long-horizon household activities, locomotion, and human-scene interaction.
• Its hierarchical benchmark reveals limitations of current methods under contact, occlusion, egomotion, and long temporal horizons.