ResNet-18 based multi-task visual inference and adaptive control for an edge-deployed autonomous robot

Current industrial robots deployed in small and medium-sized businesses (SMEs) are too complex, expensive, or dependent on external computing resources. In order to bridge this gap, we introduce an autonomous logistics robot that combines adaptive control and visual perception on a small edge computing platform. The NVIDIA Jetson Nano was equipped with a modified ResNet-18 model that allowed it to concurrently execute three tasks: object-handling zone recognition, obstacle detection, and path tracking. A lightweight rack-and-pinion mechanism enables payload lifting of up to 2 kg without external assistance. Experimental evaluation in semi-structured warehouse settings demonstrated a path tracking accuracy of 92%, obstacle avoidance success of 88%, and object handling success of 90%, with a maximum perception-to-action latency of 150 m. The system maintains stable operation for up to 3 hours on a single charge. Unlike other approaches that focus on single functions or require cloud support, our design integrates navigation, perception, and mechanical handling into a low-power, standalone solution. This highlights its potential as a practical and cost-effective automation platform for SMEs.