SpatialBot is a vision-language model designed to enhance spatial understanding by integrating RGB and depth images. The model addresses the limitations of existing vision language models (VLMs) in understanding spatial information from 2D images, which is crucial for embodied AI tasks such as robotic manipulation and navigation. To train VLMs for depth perception, the authors introduce two datasets: SpatialQA and SpatialQA-E, which include depth-related questions across various scenarios. SpatialBench is developed to evaluate VLMs' spatial understanding capabilities. Extensive experiments on the spatial-understanding benchmark, general VLM benchmarks, and embodied AI tasks demonstrate the effectiveness of SpatialBot. The model, code, and datasets are available at https://github.com/BAAI-DCAI/SpatialBot. SpatialBot uses depth information to guide VLMs in understanding spatial relationships. The model is trained on RGB-D images and depth-related questions, enabling it to perform tasks such as object detection, depth estimation, and spatial reasoning. The model also includes a depth API that allows it to query depth values of individual pixels or regions. SpatialBot is tested on various tasks, including robotic manipulation, where it successfully performs tasks such as picking up objects and placing them on specific locations. The model's performance is validated through experiments on benchmark datasets, showing significant improvements in spatial understanding and reasoning compared to existing models. The results demonstrate that SpatialBot can effectively understand depth information and perform tasks requiring spatial reasoning in both general and embodied AI scenarios. The model's contributions include the development of SpatialBot, the creation of the SpatialQA and SpatialQA-E datasets, and the design of SpatialBench for evaluating spatial understanding capabilities. The model's success in spatial understanding and reasoning highlights its potential for applications in embodied AI and robotics.SpatialBot is a vision-language model designed to enhance spatial understanding by integrating RGB and depth images. The model addresses the limitations of existing vision language models (VLMs) in understanding spatial information from 2D images, which is crucial for embodied AI tasks such as robotic manipulation and navigation. To train VLMs for depth perception, the authors introduce two datasets: SpatialQA and SpatialQA-E, which include depth-related questions across various scenarios. SpatialBench is developed to evaluate VLMs' spatial understanding capabilities. Extensive experiments on the spatial-understanding benchmark, general VLM benchmarks, and embodied AI tasks demonstrate the effectiveness of SpatialBot. The model, code, and datasets are available at https://github.com/BAAI-DCAI/SpatialBot. SpatialBot uses depth information to guide VLMs in understanding spatial relationships. The model is trained on RGB-D images and depth-related questions, enabling it to perform tasks such as object detection, depth estimation, and spatial reasoning. The model also includes a depth API that allows it to query depth values of individual pixels or regions. SpatialBot is tested on various tasks, including robotic manipulation, where it successfully performs tasks such as picking up objects and placing them on specific locations. The model's performance is validated through experiments on benchmark datasets, showing significant improvements in spatial understanding and reasoning compared to existing models. The results demonstrate that SpatialBot can effectively understand depth information and perform tasks requiring spatial reasoning in both general and embodied AI scenarios. The model's contributions include the development of SpatialBot, the creation of the SpatialQA and SpatialQA-E datasets, and the design of SpatialBench for evaluating spatial understanding capabilities. The model's success in spatial understanding and reasoning highlights its potential for applications in embodied AI and robotics.