mmET: mmWave Radar-Based Eye Tracking on Smart Glasses

With the growing popularity of VR and AR devices, eye tracking has become a critical user interface and input modality for on-device AI agents. However, a compact, power-efficient, and robust eye tracking solution for AR/smart glasses remains an unsolved challenge. In this paper, we present mmET, the first mmWave radar-based eye tracking system on glasses. Our system, implemented as a pair of prototype glasses, utilizes sub-1cm mmWave radars placed near the eyes. The radars transmit FMCW signals and capture the reflections from the eyes and surrounding skin as the system input. To refine gaze estimation accuracy and data efficiency, we propose several novel methods: (1) concatenating multiple chirps and beamforming with learnable weights to improve resolution, (2) a novel neural network architecture to enhance robustness against remounting, (3) pretraining with contrastive loss to enable fast adaptation for new users. Experiments with 16 participants show that mmET achieves an average angular gaze direction error of 1.49° within sessions and 4.47° across remounting sessions, and reduces the training data needed for new users by 80% using the pretrained model.