AdvDex: A Unified Framework for Learning Dexterous Manipulation from Human Demonstrations
A new framework called AdvDex has been developed by researchers to improve dexterous manipulation in embodied intelligence. This framework tackles the difficulties of scaling these abilities, which arise from the expensive nature of robot demonstrations and the differing action spaces in various embodiments. AdvDex utilizes demonstrations from both humans and robots, featuring two major innovations: OmniShare, a comprehensive multimodal dataset of human manipulation demonstrations that offers high-quality kinematic supervision and tactile data, minimizing the need for robot teleoperation; and the Joint-Aligned Action Space (JAAS), a standardized action representation that aligns human hand movements with those of dexterous robots. This alignment aids in cross-embodiment generalization by ensuring that task-relevant visual cues remain separate from specific embodiments. The framework is discussed in detail in a paper on arXiv (arXiv:2608.14028), which was announced as a cross-type submission. AdvDex marks a notable advancement toward scalable and generalizable dexterous manipulation in embodied AI systems.
Key facts
- AdvDex is a unified Vision-Language-Action framework for learning dexterous manipulation from human and robot demonstrations.
- OmniShare is a large-scale multimodal dataset of human manipulation demonstrations with kinematic supervision and tactile measurements.
- Joint-Aligned Action Space (JAAS) is a canonical action representation with an SE(3) wrist pose and 15 finger joints.
- JAAS aligns human hands and dexterous robot hands functionally.
- The framework aims to reduce reliance on robot teleoperation.
- It addresses cross-embodiment generalization by avoiding entanglement of visual cues with embodiment-specific appearance.
- The paper is available on arXiv with identifier 2608.14028.
- The announcement type is 'cross'.
Entities
Institutions
- arXiv