DexPort

DexPort is a portable data collection interface for high-DOF robot hands, whose core design principle is that collection and deployment use the same hand hardware.

DexPort records both commanded joint targets and measured joint states.

A small set of skills is sufficient to support a broad range of dexterous tasks, which is the cornerstone of the DexPort system design.

Using recorded robot commands rather than states raises task performance from 11% to 64%.

With in-the-wild data, task success rate rises from 36% to 78%.

Without a robot arm, the research team collected demonstration data across multiple real tasks and unstructured environments and trained visuomotor policies with behavioral cloning.

Evaluation covers four task categories: pick-and-place of diverse objects, screwing with an in-hand tool, drawer tidying requiring simultaneous grasping and pulling open, and a dual-arm scanning task.

The team used the portable device to collect over 2,500 demonstrations across laboratory and in-the-wild settings.

By having collection and deployment share the same robot hand hardware, DexPort fundamentally eliminates the hand embodiment gap, so recorded commands can be used directly for policy learning without any cross-embodiment mapping.

Whether the skill set design covers enough tasks and how the granularity of skill switching affects policy learning are not yet discussed in the paper.

A scale of 2,500 demonstrations is already substantial among real-world data, but cross-task generalization still requires more validation.