Amber Xie, Jiawei Fu, Xiaofeng Guo, Carmelo Sferrazza, Guanya Shi, Haozhi Qi
DexPort is a portable, hand-held device that enables efficient collection of dexterous manipulation demonstrations by recording both commanded joint targets and actual joint states. It allows operators to specify wrist motion and select pre-defined skills for finger coordination, reducing the need for fine-grained joint control. The system supports training visuomotor policies using recorded data, achieving high task success rates across diverse real-world tasks.
The device eliminates the need for complex robot setups during data collection, enabling high-quality demonstrations in unstructured environments. This lowers the barrier to entry for dexterous manipulation research and improves policy performance by leveraging recorded robot commands over estimated states.
Thank you to arXiv for use of its open access interoperability. arxiv.org. aibrief.ai is not affiliated with or endorsed by arXiv.