RoboMIND: Benchmark on Multi-embodiment Intelligence Normative Data for Robot Manipulation
arXiv: 2412.13877 | Venue: RSS 2025
Authors: Kun Wu et al. (37 co-authors, multiple institutions)
Abstract
RoboMIND is the largest multi-embodiment teleoperation dataset collected on a unified platform. It contains 107,000 demonstration trajectories across 479 diverse tasks involving 96 object classes, collected from four robotic platforms via human teleoperation.
Dataset Scale
- Trajectories: 107,000
- Tasks: 479
- Object classes: 96
- Embodiments: 4 platforms
- Failure demos: 5,000 labeled failure trajectories with documented causes
Hardware Platforms
| Platform | Trajectories |
|---|---|
| Franka Emika Panda (single-arm) | 52,926 |
| Tien Kung humanoid | 19,152 |
| AgileX Cobot Magic V2.0 (dual-arm) | 10,629 |
| UR-5e (single-arm) | 25,170 |
Teleoperation Method
Human teleoperation via a unified data collection platform with standardized protocols across embodiments.
Key Contributions
- Multi-embodiment on unified platform: First large-scale dataset providing cross-embodiment data under consistent collection protocols (unlike Open-X which aggregates heterogeneous data from different labs)
- Failure demonstrations: 5,000 labeled failure trajectories with cause annotations — unique among public datasets, enables learning from mistakes
- Digital twin: Paired simulation environment in Isaac Sim for sim-to-real research
- Scale: 479 tasks is ~5x more than DROID (86 tasks)
Comparison with DROID
RoboMIND focuses on task diversity (479 tasks) and multi-embodiment coverage at the cost of scene diversity (fewer unique environments). DROID prioritizes environmental diversity (564 scenes) with a single robot platform. RoboMIND adds the unique failure demo corpus.