Building the Real-World Substrate for Embodied Intelligence
High-fidelity, long-horizon egocentric datasets of human task execution in natural home environments — designed for world models, robotics manipulation, and multimodal agent research.
Why This Matters
Embodied AI and world models require training signals grounded in real human behavior over extended time horizons. Short clips and scripted demos do not capture the structure of everyday tasks, failure recovery, or the continuity of natural environments. Our datasets are built for research that needs the real-world substrate — not synthetic proxies.
Dataset Architecture
Long-horizon sessions are captured in natural home settings with head-mounted cameras. Each session is structured for temporal coherence and task focus.
Session length
3–7 hour continuous capture per participant. Minimum 15-minute segments for task-focused analysis.
Environment
Natural home environments only. Structured lighting guidelines; no staged sets or studios.
Frame content
Hands visible in frame. Manual task-focused capture. Linear FOV, stabilized head-mounted rig.
Use Cases
World models & simulation
Long-horizon, egocentric streams for learning temporal structure and environment dynamics.
Robotics manipulation
Human task execution with hands in frame; suitable for imitation and policy learning.
Multimodal agents
Vision-and-action alignment over extended sequences in real-world contexts.
Activity & intent research
Structured metadata and segment boundaries for activity recognition and planning.
Principles
- Consent-first collection. Adult-only households. Clear use cases and research context.
- Quality over volume. Manual task focus, structured guidelines, and review before release.
- Research-native. Datasets designed for frontier labs — no repackaging for consumer or creator use.
- Synchronized dual-camera sessions where needed for multi-view and 3D research.
Dataset access for research teams
Technical specs, sample metadata, and licensing are available to qualified research organizations.