Computer Vision-Based Beamforming Optimization With Trustworthy AI for 6G-IoT Services

Emerging 6G IoT frameworks and architectures operating in the terahertz band face critical challenges in near-field beamforming due to spherical wave propagation and dynamic obstacles. This paper proposes a physics-informed framework that integrates computer vision–aided monocular depth sensing with reconfigurable intelligent surface (RIS) control to enable smart, codebook-free near-field beam focusing in THz 6G IoT networks. The PINN approach weaves electromagnetic constraints directly into the learning process, ensuring beam patterns follow physical laws rather than just fitting training data. Our dual optimizer adjusts both active array weights and passive RIS elements simultaneously, responding to predicted blockages and motion detected through scene flow analysis. Tests achieved 1.07m positioning accuracy with 6.5 degree angular precision, reducing location error by 21.3 cm versus standard 5G beam management. Energy use dropped 28%, link availability rose 42%. Such improvements are vital at terahertz frequencies, where conventional methods fail and anticipatory, physics-aware control ensures reliable 6G-IoT performance.