As smart glasses and lightweight MR devices get practical, the unsolved piece is input. Controllers defeat the point, mid-air gestures are tiring and imprecise, and voice is unusable in company.
The palm is an obvious candidate, and the reason it is obvious is worth stating: it is always available, it is tactile, and you know where your own hand is without looking. That last property — proprioception — is the one that makes it different from a touchpad. You can hit a target on your own palm with your eyes closed, which is true of almost no other input surface.
What the palm does not have is a coordinate system or any touch sensing.
PalmSpace, posted 7 October 2026 by Chentao Li, Mingze Gao, Runze Sun, Zhaoguo Wang, Jianjiang Feng and Jie Zhou, is a wrist-worn infrared system that supplies both.
What it does differently
Prior on-palm work, as the authors characterise it, exposes isolated touch events, discrete regions, continuous trajectories, or task-specific gestures — each system solving one of these, none of them giving you a common representation. The consequence is that you can have precise selection or gesture manipulation, built separately, and not both through one interface.
PalmSpace jointly represents three things: whether contact is occurring, which interaction mode you are in, and palm-referenced absolute location. At the model level these coupled outputs are learned through a single shared real-time representation rather than a stack of separate classifiers.
Two properties follow, and both are harder than they sound:
Body-referenced absolute input. Not a trajectory, not a swipe — a position on the palm, in palm coordinates. The system has to establish a frame on a surface that physically deforms as the hand moves.
No per-user sensing calibration. The palm is morphologically variable — size, shape, crease pattern, skin tone all differ, and infrared reflectance is sensitive to all of them. Making it work without asking each user to run a setup routine is the difference between a research demo and something that could ship.
The numbers
Leave-one-participant-out evaluation with 17 participants — which is the right protocol here, because it tests generalisation to a hand the model has never seen rather than fit to the hands it trained on.
- 6.7 mm mean localisation error
- 98.9% contact detection accuracy
- 96.7% F1 on four-class interaction-state recognition
6.7 mm is the figure to anchor on. An average adult palm is roughly 80–90 mm across, so that error is on the order of a twelfth of the available width — enough for a grid of distinguishable targets, and comfortably inside a fingertip’s own contact patch, which is about 10 mm. The practical limit on target size here is the finger, not the sensor.
The user studies demonstrate absolute pointing and dragging, eyes-free digit input, and multi-finger controls including scrolling and pinch-based map manipulation. Eyes-free digit entry is the one that shows proprioception doing real work: you are hitting number targets on your own hand without looking at either the hand or a display.
The infrared choice
Wrist-worn infrared sits in a particular spot among the alternatives.
Cameras — head-mounted or wrist-mounted — need line of sight, fail in difficult lighting, carry a privacy cost when they are pointed at a room, and are power-hungry. EMG reads muscle activity and needs electrical skin contact, with electrode placement sensitivity to match. Capacitive sensing around the wrist works but reads tendon movement rather than palm position.
Infrared from the wrist looks across the palm from its edge, needs no skin electrodes, and consumes little power. The difficulty it inherits is the one noted above: reflectance varies with skin and with hand geometry, which is precisely why “no per-user calibration” is the claim that makes the rest of the result interesting.
Worth knowing
The name has history: a 2012 MobileHCI paper also titled PalmSpace explored around-device gestures using the palm for 3D rotation on mobile devices. Same body part, different problem — that work used the palm as a thing to be sensed in mid-air, where this one uses it as a surface to be touched.
The paper is marked “Preprint. Initial version”, filed under cs.HC with cs.CV. No code, hardware design or dataset is mentioned in the available material, and nothing is said about what happens when the hand is in motion, when the palm is partly occluded by the fingers of the other hand, or in direct sunlight — all of which are ordinary conditions for a device that is meant to be used while walking around.