Training a Robot Hand Without the Robot

X Square Robot introduced TwinDEX on 2 September from Shenzhen: a wearable data-collection device and a robotic end effector, deliberately designed as a matched pair. Both are three-finger designs with nine degrees of freedom — seven active, two passive — and, critically, they share the same kinematic chains.

That last point is the entire idea. The bottleneck in learned manipulation is not model architecture, it is demonstration data, and collecting it by teleoperating a robot is slow, requires the robot, and ties throughput to the number of robots available. Collecting it from a human wearing a device is fast and parallel, but produces data in the wrong body: a policy trained on a human hand has to be translated onto a gripper with different joints, different reach and a different camera viewpoint, and the translation is where the performance goes.

Building the collection device and the deployment gripper as the same mechanism removes that translation. The sensing is matched too — multi-view RGB camera inputs, six-degree-of-freedom wrist pose tracking, finger joint states and fingertip tactile signals, all synchronised across modalities — so the policy sees the same channels in training as in deployment.

The reported numbers support the argument. In collection evaluation, the wearable achieved up to 5.3 times the data throughput of on-robot teleoperation, and policies were trained effectively from only a few hundred robot-free episodes with no on-robot training data. A trained system executed a full 24-sub-action chemistry experiment autonomously in a single run — which is a more demanding demonstration than the usual pick-and-place clip, because a 24-step chain has no opportunity to recover from an early error. Other tasks shown include cap twisting, sweeping with a broom and dustpan, book manipulation, toolbox latch operation and syringe handling.

For industrial readers the interesting part is the scaling story rather than the dexterity. If demonstrations can be collected by ordinary staff wearing a device while doing a real task in a real environment, the cost of teaching a robot a new job drops to the cost of somebody doing the job a few hundred times — and it can happen in parallel across sites without a robot present. That is a different economic shape from programming a cell.

The claims are the company's own and the evaluation is its own; no pricing or general availability date was given. The design rationale it states rests on three things: dexterity sufficient for stable grasps and tool use, kinematic and visual consistency between collection and deployment hardware, and scalability through parallel multi-operator collection.