OmniWorld: 1B-10B Multi-Modal Records for 4D World Modeling

InternRobotics released OmniWorld in late 2025, a multi-modal collection containing between 1 billion and 10 billion records for 4D world modeling. The dataset integrates diverse domains including robotics (RH20T, DROID), human-object interaction (HOI4D), and urban environments (CityWalk).

Use Cases

• Training robotics control policies using the RH20T and DROID subsets
• Developing 3D reconstruction models from the OmniWorld-Game and HOI4D spatial data
• Generating text-to-video sequences for urban navigation using the CityWalk subset

Strengths

• Scale of 1 billion to 10 billion records
• Multi-domain coverage including RH20T robotics and CityWalk urban data
• Standardized WebDataset format for high-throughput loading

Limitations

• CC BY-NC-SA 4.0 license restricts commercial use
• Massive storage footprint required for 1B+ records
• Potential domain gaps between synthetic game data and real-world robotics sensors

Provenance

Source

InternRobotics (ICLR 2026)

Collection Method

Aggregated from multiple primary sources including robotics sensors, human-object interaction captures, and synthetic game environments.

Time Range

2025-2026

Freshness

Last updated March 2026.

Geography

Global (includes urban walks and robotics lab environments)