Emerging Interfaces

GazeTags Let Smart Glasses Know What You're Looking At — Without Ever Using a Camera

CMU's Future Interfaces Group built $2 infrared tags that only respond when a narrow, head-aligned beam from a wearable hits them — giving object-aware AI a camera-free, privacy-preserving, year-long-battery path.

Giving a wearable device awareness of what its user is looking at usually means a camera and a computer-vision pipeline running constantly — expensive on battery, and an obvious privacy liability for anything worn on your face all day. GazeTags, a UIST 2026 paper from Carnegie Mellon’s Future Interfaces Group, solves the same problem without a camera at all.

How it works

A user-worn device — smartglasses, hearables, anything with the hardware — periodically emits a narrow, focused infrared beam aligned with the wearer’s head direction. Small, cheap electronic tags placed on objects around a home or office sit passive until that beam actually hits them; only a tag caught in the beam wakes up and replies with its identity and, optionally, a live payload like sensor state. Everything outside the beam’s narrow cone simply never responds, which is what makes the whole approach camera-free: there’s no image being captured or analyzed anywhere in the loop, just a directional radio handshake.

Each tag costs roughly $2 to build and runs for more than a year on a single coin-cell battery — a direct consequence of staying passive until addressed, rather than continuously broadcasting or continuously imaging. In testing with five tags placed close together, the system identified all of them correctly in 99.75% of trials at distances of four and six meters; accuracy dropped at longer range, particularly as more tags were added to a cluster.

The tradeoffs, stated plainly

GazeTags tells you head direction, not eye direction — a real gap from gaze-tracking systems that follow where your eyes actually point rather than where your face happens to be aimed. And every object you want the system aware of needs a tag installed beforehand; this isn’t a general-purpose scene-understanding system, it’s an awareness layer for a pre-instrumented environment. Both are reasonable costs for what the approach buys in return: continuous operation on battery-constrained wearables, and a privacy model that doesn’t depend on trusting a company’s camera-data-handling policy, because there’s no camera data to handle.