DexPort Records the Commands Behind a Robot’s Grip

In one of the DexPort team’s tests, a robot learned to pick up a package but struggled to keep hold of it long enough to scan its barcode. The problem was not simply that it had seen too few examples. It had been taught to imitate where the fingers ended up, rather than the commands that kept them gripping.

The researchers changed that training signal while keeping the rest of the learning setup the same. In their small barcode-scanning evaluation, completion rose from zero to 73%. Each version was tested for 15 attempts, so this is evidence about a particular experiment, not a warehouse reliability rating.

DexPort, a preprint posted on October 23, 2026, describes a portable way to collect that missing information. A person carries and operates an actual robot hand, recording its instructions alongside its movements. The same hand-and-camera module later attaches to a robot arm. The idea is to make demonstrations easier to gather without discarding the details that turn a plausible-looking movement into a working grip.

A Finger’s Position Does Not Tell the Whole Story

A robot’s instruction to a finger and the finger’s actual position are two different things. During contact, an object can prevent the finger from reaching the position it has been told to reach. That difference matters: copying the final position alone does not preserve the command used to hold the object.

DexPort records both. Its “commanded joint targets” are the requested positions sent to the hand’s motors. Its measured joint states describe where the joints actually are. The researchers’ comparison trains one version on the commands and another on the measured positions, using the same remaining training recipe.

The position-trained version often produced weak grasps. It could drop the package during scanning or lose a piece of laundry while pulling a drawer open. These failures make the distinction tangible. A demonstration can show an object held in place without fully specifying the action that holds it there.

This is a different problem from labeling what happened in a robot video. A useful video label might identify the attempted action and whether it succeeded. DexPort captures commands from the robot hand itself. Those sources of training information can serve different purposes; the study does not show that video-based learning is useless.

Take the Hand, Leave the Robot Arm

Gathering demonstrations through a complete robot usually ties the operator to that setup. DexPort separates the hand from the arm during collection.

The device combines a five-fingered Wuji robot hand, a wide-angle camera and a handle built around a Meta Quest 2 controller. The person moves it through the task and uses a few controls to coordinate the fingers: a whole-hand grasp, a pinch, a two-finger grasp or a learned tool-rotation skill. There is no need to direct each of the hand’s 20 movable joints individually.

Those shortcuts help the human demonstrator. The resulting robot policy, the learned system that chooses actions, is trained on the recorded hand commands and wrist movements rather than on labels for those shortcuts.

Reusing the same physical hand and camera also avoids translating a human hand’s shape and movements onto different robot fingers. It does not eliminate every setup step: the wrist movements still need calibration to the robot arm’s coordinate system.

The team collected more than 2,500 demonstrations across laboratory, apartment and office settings. Portability has a physical cost, though. The assembly weighs about 955 grams, and the authors say its weight limits uninterrupted use and rotation-heavy movements. This is a tool someone carries, not frictionless data collection.

Better Grips, Not Better at Everything

The main comparison covered placing vegetables on a plate, putting an object into a drawer, scanning a barcode with two hands and turning a screwdriver within the hand.

Across the three tasks scored by completion rate, the authors report an average of 64% with recorded commands versus 11% with measured positions as the training targets. That average excludes the screwdriver experiment, which measured how long the tool stayed in the hand and how quickly it turned. Each task used 15 evaluation attempts per version.

The screwdriver result resists a simple “everything improved” story. Command-based training turned the tool faster, about 3.8 rotations per minute versus 1.2. But it held the tool for less time: 15.6 seconds versus 31.7. Keeping an object and doing useful work with it are not identical achievements.

Likewise, a secure grip is not necessarily a gentle one. DexPort’s experiments support a claim about the training information needed for these tasks, not a general guarantee that a robot can handle fragile goods safely.

More Backgrounds Do Not Fix Every Failure

A portable teaching device is attractive partly because it can collect examples away from a fixed laboratory table. With 700 laboratory demonstrations as the base, adding 1,000 examples from apartments and offices raised success on unfamiliar objects from 20% to 73% in the pick-and-place test.

The extra demonstrations used the same training objects in different surroundings. They varied backgrounds and lighting rather than introducing a thousand new object shapes.

The appendix supplies an important counterexample. Starting with only 500 laboratory demonstrations, adding those 1,000 outside examples reduced success on unfamiliar objects from 20% to 7%. The authors suggest that a sufficiently large laboratory dataset may be needed for the benefits of scene variety to transfer to new objects. These small tests do not establish a universal data threshold.

DexPort therefore offers more than a convenient way to collect robot videos, but less than a solved recipe for teaching any robot hand. Its clearest lesson is about what gets preserved in a demonstration. Watching where a finger stops can miss the instruction that makes it hold on.