PaXini’s new glove feels 0.1N of force to teach robots delicate work

Teaching a robot to tie a delicate ribbon or handle a balloon is usually an exercise in frustration. While robot hands are getting mechanically impressive, they still struggle with delicate tasks because they have no real sense of human touch or touch pressure.

PaXini AI introduced the PXCap Pro, a new five-finger data collection glove designed specifically to fix that bottleneck. Instead of guessing how humans interact with physical objects, the glove captures delicate human demonstrations and turns them directly into training data for embodied AI models.

Touch, vision, and joint tracking in one package

Traditional data collection usually relies on external cameras or clunky mocap setups that miss critical contact forces. The PXCap Pro bundles tactile sensing, wrist vision, and joint tracking into a single synchronized hardware system.

Across the fingertips and palm, multi-dimensional array tactile sensors can detect force changes down to 0.1 Newtons. That level of sensitivity lets it register tiny shifts in three-dimensional and distributed pressure that conventional setups miss entirely.

On the wrist, a 165-degree ultra-wide-angle camera tracks hand movements and object interactions up close, eliminating the blind spots common with fixed room cameras.

To capture finger posture, PaXini built in angular encoders with an accuracy of ±0.15 degrees. Crucially, they are engineered to resist electromagnetic interference, keeping joint angle tracking rock-solid even around running machinery.

The entire glove weighs just 300 grams and fits close to natural hand dimensions. It comes pre-calibrated from the factory with permanent calibration, and onboard processing streams clean, structured data straight out of the box, ready for neural network training without post-processing headaches.

Putting gloves on workers to train machines

In practice, workflow collection looks refreshingly straightforward. Operators or technicians put on the glove and simply perform delicate tasks as they normally would.

Those tasks can range from packing shipping boxes and tying gift ribbons to handling fragile materials like balloons or operating pipettes in a laboratory. Over time, industries like automotive assembly, food prep, and eldercare could use this to document expert human workflows.

Historically, data collected across different operators suffered from mismatched hand sizes, inconsistent sensors, and noisy readings, leaving teams with isolated, unrepeatable datasets. PaXini is pushing a standardized hardware baseline so companies can build reusable physical interaction datasets. The company plans to showcase the glove at the 2026 World Robot Conference starting August 19.

Why it matters

For roboticists, the PXCap Pro eliminates months of custom rigging. Trying to synchronize third-party tactile skins, motion gloves, and external vision usually results in a messy latency nightmare. Having a 300-gram unit that streams calibrated tactile and visual ground truth out of the box makes gathering real-world training runs far more practical.

Across the industry, hardware form factors for humanoids and robotic arms are quickly converging. The real battleground has shifted to the foundation models driving them, where high-fidelity physical interaction data remains the ultimate bottleneck. Focusing on the pick-and-shovel infrastructure of tactile data collection looks like a very sharp move by PaXini.

The main friction point will be cross-embodiment retargeting. Capturing exact human hand movements and 0.1N forces is one thing, but translating those signals reliably onto diverse robotic grippers and multi-finger hands with different kinematics will still require serious algorithmic heavy lifting on the model side.