DexPort: A Hand-Held Skill Interface That Closes the Dexterous Data Gap
DexPort is a portable collection interface for high-DoF robot hands that uses the same hand hardware for demonstration and deployment, removing the hand embodiment gap entirely. The operator moves a hand-held interface to specify wrist motion and composes a small library of skills to coordinate the fingers, instead of driving every joint. The device records both commanded joint targets and measured joint states. Across four real tasks and more than 2,500 demonstrations collected in the lab and in the wild, behaviour-cloned visuomotor policies improve from 11% to 64% success when trained on commands rather than states, and from 36% to 78% when in-the-wild data is added.
Paper · arXiv:2610.32969Dexterous manipulation is bottlenecked by demonstrations, not by algorithms. Teleoperation logs actions a robot can execute directly, but it presupposes a complete robot and loads the entire burden of finger coordination onto the operator. Human demonstrations are cheap to gather, yet they must be estimated and retargeted before a robot can use them, and a hand embodiment gap survives the retargeting. Both routes are expensive in exactly the dimension that matters: scale.
Core innovation
DexPort’s move is to make the collection rig and the deployment hardware the same device. The hand-held interface houses the target dexterous hand itself, so every recorded demonstration already lives in the robot’s own kinematic and sensor space. No human-to-robot mapping is learned, because none is needed. The system logs two signals per timestep: the commanded joint targets and the measured joint states.
The interaction design matters as much as the hardware. The authors’ key observation is that a small set of skills supports a wide range of dexterous tasks. During collection the operator only translates the interface to specify wrist motion, while finger motion is produced by selecting or adjusting a skill. Precise grasping and in-hand manipulation are preserved, the operator is freed from per-joint control, and the resulting data trains a visuomotor policy by plain behaviour cloning with no skill labels attached.
Results
- Four real-world tasks: pick-and-place over diverse objects, screwing with in-hand tool use, drawer stowage requiring simultaneous grasp and pull, and a bimanual scanning task.
- Over 2,500 demonstrations collected with the portable device, in laboratory and in-the-wild settings, the latter without a robot arm present.
- Supervising on recorded robot commands instead of joint states lifts success from 11% to 64%.
- Adding in-the-wild data lifts task success from 36% to 78%.
Limitations
The skill library is still hand-designed, so the policy’s ceiling is set by what the skill set covers. Generalisation is demonstrated on four task families only; cross-embodiment transfer and long-horizon tasks are not evaluated. Collecting without a robot arm lowers the barrier, but the interface’s own pose still needs external tracking, and the sensitivity of demonstration quality to tracking error is not quantified.
Industry implications
What dexterous-hand vendors lack is not another hardware revision but evidence that the hardware is usable at scale. DexPort reframes collection from “build a teleoperation station” to “pick it up and record”, and because the collecting hand is the deploying hand, the usual performance collapse after a hardware swap never occurs. For teams building hands, manipulation policies, or data services, this is a collection pipeline that can be copied almost verbatim. For full-stack robot makers, it means hand data can scale distributively, like field data collection, instead of consuming scarce hours on the robot itself.