RealSource World provides over 14 million frames and 11,428 episodes of real-world robotics manipulation data. Collected using an RS-02 dual-arm humanoid robot, it covers 35 distinct long-horizon manipulation tasks with annotations for atomic skills and quality assessments.
Use Cases
- Train imitation learning models on the 14+ million frames of dual-arm manipulation demonstrations.
- Analyze the sequence and success of atomic skills within the 11,428+ annotated episodes.
- Benchmark robotic policy performance across the 35 distinct manipulation task categories.
- Develop skill segmentation algorithms using the detailed annotations for atomic actions.
Strengths
- Contains over 14 million frames of real-world video data.
- Includes 11,428+ demonstration episodes across 35 distinct tasks.
- Features detailed annotations for atomic skills and quality assessments.
Limitations
- Dataset size, file formats, and specific column structure are not detailed in the available input.
- The temporal and geographic scope of the data collection is unspecified.
Provenance
- Source
- RealSourceData on Hugging Face.
- Collection Method
- Collected using an RS-02 dual-arm humanoid robot performing tasks in real-world environments.
- Time Range
- null
- Freshness
- Last updated in February 2026.
- Geography
- null