ARUNet: Advancing Real-Time Stereo Matching for Robotic Perception on Edge Devices

Accurate real-time 3D depth perception is crucial for robotic systems, enabling navigation, obstacle avoidance, and object manipulation. However, the high computational demands of stereo matching networks hinder their application on resource-constrained robotic platforms. This paper presents ARUNet, an efficient stereo matching network optimized for robotic perception on edge devices. We introduce an adaptive multi-scale feature extractor (AMFE) for fast extraction of high-precision features, a hierarchical group-wise distance method to improve cost volume similarity and a three-stage aggregation strategy for cost volume refinement. The first stage uses 3D convolution for global regularization, followed by depthwise separable convolutions and residual connections in the subsequent stages for efficient refinement. A hybrid disparity upsampling module (HDUM) combining bilinear and nearest-neighbor techniques is proposed to maintain disparity clarity while reducing computational cost. ARUNet balances speed and high-precision disparity estimation on the NVIDIA Jetson AGX Orin, making it suitable for robotic perception. Experimental results on the KITTI 2012 and KITTI 2015 datasets show that ARUNet achieves state-of-the-art performance on resource-constrained edge devices.