Real-time vision-based obstacle avoidance for mobile robots using lightweight monocular depth estimation and behavior-driven control

Safe and real-time navigation of mobile robots without relying on costly sensors remains a major challenge in robotics and intelligent systems. This study proposes a lightweight, low-cost, and noise-resilient framework that integrates monocular depth estimation (MDE) with behavior-based control to achieve obstacle avoidance and autonomous motion without GPU or LiDAR. The proposed MDE network operates at 14–20 Hz on embedded CPU hardware, achieving a mean absolute depth error (MAE) of 0.056 m, root-mean-square error (RMSE) of 0.082 m, scale-invariant logarithmic error (SILog) of 0.017, and relative error of 3.9%. Field experiments in unstructured indoor and outdoor environments demonstrate consistent navigation success rates of 96.7%, path efficiency of 84.6%, and average course-completion times of 3.8 ± 0.4 min for complex scenarios with 10–15 obstacles. The computational load remains under 68% CPU utilization with memory usage below 450 MB, ensuring stable operation on low-power platforms such as Raspberry Pi 4 and Jetson Nano (in CPU mode). The proposed system introduces a computation-efficient MDE network for real-time, GPU-free inference, a 50 Hz ROS-integrated control loop for smooth motion planning, and comprehensive real-world validation, delivering a practical, cost-effective navigation solution for service, rescue, and industrial robotics under constrained resource conditions.