Smart glasses, AI pins and earbuds keep running into the same problem. To be useful, the device has to know what in the room you care about right now. The usual answer is an outward-facing camera feeding a computer-vision model. That costs power and compute, and it puts a lens in front of everyone the wearer looks at. Researchers at Carnegie Mellon University's Future Interfaces Group have a much simpler answer: give the objects a cheap infrared tag, and have the wearable ask whichever tag it is pointed at.
The system is called GazeTags. It is described in the paper Camera-Free Object Discovery and Interaction with GazeTags by Vimal Mollyn and Chris Harrison, published in the Proceedings of UIST '26 as Article 12, pages 1–10 (doi 10.1145/3830398.3830713), according to the lab's project page. The group describes it as head-gaze object awareness that needs no camera and no computer vision. The lab says the tags cost roughly $2 each and run for more than a year on a coin cell.
This is a research project. Neither the lab page nor the coverage we found mentions a product, a kit, pre-orders or a release date.
How it works
The project page explains the mechanism. The wearable periodically sends out a focused infrared signal lined up with the direction of the wearer's head. Only tags inside that beam wake up. Each one replies with its identity and, if it has them, extra data such as the object's state or a live sensor reading.
That asymmetry is the key design choice. The tags sleep most of the time and wake briefly to listen for incoming signals, per Hackster, so they can spend nearly all of their life in a low-power state. The wearable doesn't have to understand the scene, because the beam does the selection: whatever sits in the IR cone is, in effect, what the user is facing.
Chris Harrison's research page frames the goal in similar terms. Wearables need to know what in the environment is relevant, and GazeTags target a coarser head-vector notion of gaze rather than true eye tracking.
The hardware, as reported
The prototype hardware details come from Hackster.io's write-up. We could not read the paper PDF or the ACM Digital Library entry, so we have not checked these figures against the paper itself.
Wearables. The team built both glasses and earbuds. Per Hackster, each uses:
- An ESP8285 SoC
- An IR receiver
- A 940 nm IR LED
- A beam roughly 40 degrees wide
Tags. Hackster describes each tag as:
- 30 mm across and 18 mm thick (Harrison's page also gives 30 mm in diameter and 18 mm thick)
- Built around an ATtiny85 microcontroller
- Fitted with an IR transmitter and an IR receiver
- Powered by a CR2477 coin cell
- Rated for an estimated battery life of about 1.4 years
The 1.4-year estimate fits the lab's own claim of "more than a year" on a coin cell. Neither source we read gives clock speeds, the IR modulation scheme, the data rate or current draw, so we are not quoting any. The roughly $2 figure is the lab's per-tag cost estimate, which Harrison's page says includes the battery. It is not a retail price, and no price is given for the wearables.
What the tags can say
The reply payload is what makes this more than an IR remote run backwards. Hackster's examples:
- A plant that reports its soil moisture
- A kettle that reports its temperature
- A fire extinguisher that reports its service date
Each is a small piece of context that a voice assistant or heads-up display could use the moment you look at the object. No image is captured, and nothing has to be recognized.
Accuracy and limits
The headline number, per Hackster, is that five closely grouped tags were correctly identified in 99.75 percent of trials at both 4 m and 6 m. Hackster reports that performance dropped at longer distances, especially as more tags were added, but gives no figures for those distances, so neither do we. Harrison's page says only that the system works at ranges up to 8 m and with many tags operating at once. "Closely grouped" is the hard case here: with a beam about 40 degrees wide, telling neighboring tags apart is the test of whether head direction alone is precise enough.
The limitations are stated plainly in the coverage:
- Head direction, not gaze. The system knows where your head points, not where your eyes look. If you glance sideways without turning your head, it won't follow.
- Pre-tagged objects only. GazeTags only knows about things someone has already tagged. Unlike a vision model, it cannot spot an object it has never seen.
- Shorter range is better. The reported 99.75 percent figure applies at 4 m and 6 m, and Hackster says performance dropped at longer distances.
Those are real trade-offs, but they are the cost of a system with no camera, no image processing and coin-cell power at the tag end. For a fixed set of objects you care about, like plants, appliances or safety equipment, pre-tagging could be a one-time job rather than a constant limitation.
What It Means for Makers
Who it's for. Right now GazeTags is a research prototype, aimed at the HCI research community. Makers working on DIY smart glasses, assistive devices, or "look at it to control it" home automation may find the architecture a useful reference.
What it replaces or enables. The approach stands in for camera-plus-vision pipelines when the job is "which of my known objects am I facing?" rather than open-ended scene understanding. Here the directional IR beam does the selecting, and tags stay quiet until they are addressed. Privacy is a side benefit the lab itself highlights: a wearable with no outward camera sidesteps many of the concerns around image capture.
Familiar parts. The parts named in the reporting, including the ESP8285 SoC, the ATtiny85, a 940 nm IR LED, an IR receiver and a CR2477 cell, are all off-the-shelf components. The sources describe the architecture but not the protocol, firmware, or whether design files will be released. Anyone hoping to reproduce it should wait for the paper or any materials the authors publish rather than guess at the details.
Price and availability. The Future Interfaces Group puts the tag cost at roughly $2 USD each (lab estimate, per the project page, November 2026). GazeTags is a research prototype presented at UIST 2026. The sources we read mention no commercial availability, kit, pre-order, pricing for the wearables, or open-source release.
A note on sourcing. Two of our three sources are the creators' own pages. The only independent coverage we found is Hackster.io, which supplied the part numbers, beam angle, battery estimate and accuracy figures. We could not access the paper itself, so treat the hardware details as reported rather than independently confirmed.