Researchers have developed a new AI model called WNM-3D for vision-language navigation. This model uses 3D scene conditioning to improve the performance of continuous navigation tasks. It does this by extracting geometry-aware representations from past observations and using them to condition future predictions. The model is trained through a combination of supervised learning, imitation learning, and reinforcement learning. Experiments show that WNM-3D outperforms other models in closed-loop navigation tasks. --- Why it matters: This matters because it addresses the limitations of existing vision-language models in continuous navigation tasks. By incorporating geometry-aware representations, WNM-3D can better model how an agent's visual observations should evolve under its predicted motion, leading to improved performance in real-world scenarios. Source: https://arxiv.org/abs/2608.07267