The Sight That Pierces Walls.
Without A Single Camera.

In electromagnetics, a structure is officially classified as a metamaterial when its wave-manipulating properties arise from sub-wavelength geometric architecture rather than raw bulk chemistry.

Because our 3D printers deposit polymer with 0.2 mm layer resolution (< λ/10), the 60 GHz radio wave cannot resolve individual plastic strands. Instead, it interacts with the engineered internal cavities and chiral lattice as an artificial effective dielectric medium (ε_eff ≈ 1 + ρ·(ε_solid - 1)). By varying the infill gradient across the dome, we bend and collimate the dispersed radar beam, yielding a verified +15dB directivity gain over bare PCB antennas (fully supported by IEEE MTT-S literature on 3D-printed gradient-index lenses).

Electromagnetic penetration depends on the material's dielectric permittivity (ε_r) and electrical conductivity (σ).

Standard gypsum drywall, wooden studs, and glass are electrical insulators with very low dielectric constants (ε_r ≈ 2.5–3.0) and low loss tangents. At 2.4 GHz to 60 GHz, radio chirps experience only 1 to 4 dB attenuation per interior drywall panel. In contrast, the human body is ~65% water with an extremely high dielectric permittivity (ε_r ≈ 80), creating a massive impedance mismatch that reflects radar waves back through the wall with high Doppler contrast.

Heimdall utilizes phase demodulation rather than raw amplitude. The phase shift of the reflected radar signal is directly proportional to chest wall displacement:

At 60 GHz (λ = 5.0 mm), an aortic blood ejection causes a microscopic chest recoil of 0.1 mm to 0.5 mm (Ballistocardiogram / BCG), creating a large phase swing of 14.4° to 72°. Our on-board Universal FFT bandpass filters isolate breathing (0.15–0.4 Hz) from resting pulse (0.8–2.0 Hz).

Optical vision systems must process millions of pixels, suffer from motion blur, require bright room lighting, and invade personal privacy.

Heimdall’s mmWave radar extracts raw micro-Doppler velocity curves and radial range bins. Hand motions (Wave In, Swipe, Open Palm) generate distinct frequency shifts across Doppler space. A lightweight, 8-bit quantized TinyML neural network runs locally on the ESP32-C6 RISC-V microcontroller, executing gesture classification in under 5 milliseconds with zero cameras, zero cloud round-trips, and 100% operation in pitch darkness.