Stereo cameras imitate how human eyes, perceive depth information in their surrounding environments. This is crucial for tasks in robotics, especially tasks for Automated Guide Vehicle (AGV) such as localization, navigation, and mapping. Existing studies based on convolutional neural networks (CNN) have gained promising results on depth estimation. However, AGV encountered difficulty applying previous studies due to their large computational time, leading to low frames per second. Therefore, we propose a lightweight stereo network for AGV and deploy it in realistic environments on Jetson Nano and a stereo camera. It consists of a lightweight backbone for feature extraction, a 3D deep encoder-decoder architecture, and an up-sampling module for disparity estimations. To improve computational time, the networks calculate a disparity map at low resolution and then refine the disparity maps at higher resolution. Through evaluation, the proposed method enhances accuracy while the archive promises computational time compared to real-time stereo networks.