UNIST Develops LightSplat: Real-Time 3D Spatial AI Understanding Human Language

Researchers at the Ulsan National Institute of Science and Technology (UNIST) have developed LightSplat, an open-vocabulary-based 3D spatial recognition technology that allows robots and augmented reality systems to locate objects using natural language queries. LightSplat uses 2-byte indices instead of high-dimensional data for each 3D point, dramatically reducing memory usage to 1/64th of existing technology. The semantic injection time was reduced to approximately 5 seconds (50-400x faster than state-of-the-art solutions). Performance metrics include: Memory usage reduced to 1/64th; Semantic injection time of 5 seconds (vs. 4 minutes to 100 minutes); Inference time of 0.002 seconds per query; mIoU score of 37.11 across 19 categories on ScanNet dataset. The technology enables robots to understand verbal commands and precisely locate objects in 3D space. Applications include robotics, AR/VR content creation, and digital twin technology for industrial environments. The research was accepted at CVPR 2026.