Abstract

The increasing demand for real-time data processing in Internet of Things (IoT) systems stems from the need for responsive and efficient services across a wide range of applications. A clear example is Human-Robot Interaction (HRI), where service robots-operating as advanced IoT nodes-face constraints in processing power and battery life. To address these limitations, many systems have adopted cloud platforms and remote processing services, enabling robots to offload computation and respond more effectively. However, this architecture introduces a significant challenge: data transfer time. High transfer latency can hinder interaction quality and degrade overall system performance. Although load-balancing strategies have been implemented to distribute tasks across servers, most do not consider the influence of transfer latency on decision-making. To fill this gap, this work proposes a machine learning-based model to predict data transfer time using features such as IP addresses, data size, geographic distance, date, and request time. Among the models tested, the Random Forest algorithm showed the best performance, achieving an R-squared value near 1.0 and the lowest errors in MSE (44.3188), MAE (1.9068), and RMSE (6.6572). A web-based application was developed to implement and validate the model in real time. This tool can be integrated into existing load-balancing systems, allowing them to consider predicted latency when distributing computational tasks. The proposed solution thus enhances resource management in HRI environments and contributes to more autonomous, adaptive, and latency-aware infrastructures across the broader IoT landscape. © 2025 IEEE.