Summary
DexVerse is a large-scale modular benchmark covering 100 dexterous manipulation tasks spanning grasping, relocation, articulated-object interaction, tool use, bimanual coordination, nonprehensile control, multi-goal execution, and long-horizon multi-stage completion. It supports 3 robot arms and 6 dexterous hands, with configurable visual variations (texture, background, lighting, viewpoint) for visuomotor generalization evaluation.
Key Contributions
- 100 diverse tasks covering the full breadth of dexterous manipulation interaction modes
- 3 robot arms × 6 dexterous hands for multi-embodiment evaluation
- 3,180 expert demonstrations with synchronized proprioceptive, RGB, depth, point-cloud, and state observations collected via VR teleoperation
- Configurable visual variation for systematic sim-to-real and cross-domain generalization assessment
- Extensible design: new tasks, assets, and embodiments can be added via modular task API
Significance
Addresses a critical gap in the field — prior benchmarks remain limited in task diversity, embodiment coverage, or controllable visual variation — providing a comprehensive and reproducible evaluation platform for the next generation of dexterous robotic policies.