ACE-Data-0: 150-Hour Multimodal Dataset for Embodied AI

Yukang Cao · hf · 2026-07-31

To address the data bottleneck in embodied intelligence, researchers introduce ACE-Data-0, a dataset designed to capture the full perception-action loop, including egocentric views, whole-body motion, dexterous manipulation, and multisensory signals over time.

Data Engine & Collection:

Uses the Ambient Capture Engine (ACE) to transform real home environments into spatially calibrated, synchronized recording studios.
Operates at table-scale (hand-object manipulation) and room-scale (whole-body locomotion).
Records egocentric/exocentric video, full-body and hand motion, object 6-DoF trajectories, audio, and tactile signals as a unified stream.

Scale & Evaluation:

Comprises 150 hours and 17M frames across 200 task categories, with 75,000 interaction episodes from 50 participants.
Covers atomic manipulations, long-horizon household activities, and human-scene interactions.
Introduces a hierarchical benchmark revealing that state-of-the-art methods struggle significantly with contact, occlusion, and long temporal horizons.

Original post →

ACE-Data-0: 150-Hour Multimodal Dataset for Embodied AI

More from Embodied