A portable tool for collecting data for robot hands
The DexPort interface uses hand hardware to record motions for dexterous manipulation tasks.
Researchers introduced DexPort, a portable data-collection interface for high-DoF robot hands. High-DoF refers to many degrees of freedom. This device is used for collecting demonstrations for learning dexterous manipulation.
The device uses the same hand hardware for collection and deployment. This avoids a hand embodiment gap. An embodiment gap occurs when the collection tool and the robot differ.
An operator moves the hand-held interface to specify wrist motion. They select or adjust skills to coordinate finger motion. Skills are a small set of motions that support many dexterous tasks.
The system records both commanded joint targets and measured joint states. Researchers used behavior cloning to train visuomotor policies. These policies are trained on recorded wrist motions and hand commands without skill labels.
The device was used for 4 tasks including pick-and-place and screwing with in-hand tool use. It was also used for drawer stowage and a bimanual scanning task. Over 2,500 demonstrations were collected in laboratory and in-the-wild settings.
Task success improved from 11% to 64% when using recorded robot commands instead of states. Success also improved from 36% to 78% when using in-the-wild data.
Key facts
- DexPort is a portable interface for high-DoF robot hands.
- The system recorded over 2,500 demonstrations.
- Task success reached 78% when using in-the-wild data.