10Kh RealOmin: 1 Million Robotic Interaction Clips for Embodied AI

10,000+ hours of cumulative robotic interaction data and over 1 million video clips for embodied intelligence research, released by genrobot2025 in early 2026. The collection features dual-arm robotic interactions with daily objects, including 35,000 manually sorted clips from the Stage 2 update. It focuses on real-world scenarios involving flexible, irregular, and various-sized objects in diverse storage environments.

Use Cases

• Training Vision-Language-Action (VLA) models using the 1 million+ video clips and dual-arm interaction data
• Developing reinforcement learning policies for robotic manipulation of flexible and irregular objects
• Fine-tuning embodied AI agents for manual sorting tasks in domestic or industrial scenarios

Strengths

• 10,000+ hours of cumulative robotic interaction data
• 1 million+ video clips
• Manual sorting and organization of daily objects
• Enhanced FOV for lower environment visibility

Limitations

• Potential geographic bias toward the US region as indicated by metadata tags
• High storage and compute requirements due to the 10,000-hour video scale
• Lack of specific column-level documentation for telemetry data

Provenance

Source

genrobot2025

Collection Method

Real-world robotic sensor capture and manual annotation/sorting

Freshness

Updated February 2026 with Stage 2 data additions.

Geography

United States