Abstract
<title>Abstract</title> <p>As artificial intelligence (AI) moves from digital interpretation to embodied action, robotic perception must move beyond visible-scene description towards physical representations. Inspired by dragonfly compound eyes, Dragonfly-TeraVision is proposed as a task-driven heterogeneous sensing architecture for physically interpretable environment reconstruction. The system integrates panoramic vision for dense semantics, Light Detection and Ranging (LiDAR) for metric line-of-sight (LoS) geometry and monostatic terahertz (THz) sensing for propagation-aware structure and material-sensitive reflection responses. A task-driven agent decomposes each sensing objective, selects modality-specific processing modules and associates their outputs through physics-guided multi-modal constraints. As a demonstration, Dragonfly-TeraVision reconstructs semantic three-dimensional (3D) geometry in an indoor environment, by detecting THz-derived hidden structural evidence beyond optical/LiDAR visibility and associating semantic surfaces with material-sensitive THz responses. This work establishes a route towards environment reconstruction for embodied sensing systems that links visual semantics, metric geometry and material-sensitive THz propagation and reflection responses within a unified physical scene representation.</p>