Wearable tactile gripper: you won't know if it's worth it until you try it on

A friend recommended the Qianjue XTac UMI G1 set to me, saying their team has been using it recently and that collecting tactile data with it is way less hassle than dragging a real robot around. Back in August at the World Robot Conference, I got to touch their sensor at the booth — pressing down showed real-time 3D deformation and force distribution. That time my impression stopped at the level of "it can feel slippage." This time my friend lent me the device and I ran it for an afternoon, so I went from the booth to actual output.

First, what this thing is. The XTac UMI G1 is a wearable visuotactile multimodal data collection gripper set, and it breaks down into five parts. The master gripper, the one a person wears on their hand to make motions; the slave gripper, mounted on a robot to replicate the motions; a collection backpack, with the compute and storage inside; a headset kit; and a visualization terminal to see what's being collected in real time. It collects tactile signals, plus RGB images, visuotactile images, 6-DoF pose trajectories (three directions of position plus three directions of orientation), gripper state, and IMU — inertial measurement unit, the chip that records motion and orientation. The official specs are 5 ms-level multi-device time synchronization and 3 mm-level spatial positioning accuracy. I can't independently reproduce those two numbers, so I'll just say they're the claimed specs.

It sidesteps an old problem.

Previously, collecting robot manipulation data basically came down to two paths. One was teleoperation — a person uses a controller or motion-capture glove to drive a real robot, and data is read from the real robot's joints. The upside is native compatibility with the real robot; the downside is you need a real robot sitting there, and one isn't enough, you need several, so cost and time both pile up. The other was pure vision — a few cameras filming the human hand. Cheap, but the tactile layer is empty. For working conditions like flexible materials and precision small parts, you simply can't judge how much force to use just by looking.

XTac takes a third path: a person wears the master gripper to make motions, the slave gripper replicates in sync, and tactile signals are collected directly by the sensor. The upside is that data collection doesn't depend on the real robot itself — one device can be taken to different scenarios, lowering the cost barrier. The key at this layer is whether the collected data can be fed directly into real-robot training with minimal transfer in between; the hardware itself isn't actually the point.

There are sticking points too. The calibration step was slower than I expected — the visuotactile sensor has to be aligned with the pose system, and the first time took about twenty minutes. My friend said with practice it can be compressed to under ten minutes, but you still have to redo it every time you change scenarios. Wearing it is constraining — carrying the backpack around for an afternoon puts a load on your shoulders. Also, the master gripper's size is made for common hand shapes, so people with small hands find it awkward to do fine motions — I tried pinching small parts a few times and kept missing.

The visualization terminal was a pleasant surprise. On screen you can see the RGB feed and the tactile map at the same time; when pressing and sliding, the force distribution and slip state change in real time. That feeling of "seeing the force" is pretty intuitive. I collected one set of insertion/removal and one set of grasping; after the data was written to disk, the structure was more organized than I expected, which saved a lot of cleanup work later.

Conclusion here, depends on the situation. For teams doing embodied algorithms that need tactile data and don't want to be bottlenecked by the number of real robots, this set is worth putting in the budget for a trial round; for scenarios that need large-scale collection like data factories and university labs, its cost-effectiveness will show up even more clearly. But if you're only doing vision-based imitation learning, or your team is so small you collect data only a few times a year, the time cost of calibration and wearing isn't worth it — not recommended. Individual developers, forget it even more — this isn't something you can just take home and play with.

To judge whether you should touch this, it really comes down to one sentence: are you short on tactile data, or short on data? If the former, this device is currently one of the few paths that sidesteps the real robot. If the latter, sort out your collection workflow first.