WalkGPT provides grounded vision-language conversation with depth-aware segmentation

The model addresses limitations of large vision-language models in reasoning about spatial aspects of urban scenes for pedestrian navigation. It uses depth-aware segmentation to ground conversations in real-world geometry.