Adaptive Velocity Control Based on IMU Terrain Classification for Campus Delivery Robots

Abstract

The ability of autonomous robots to adapt to a wide range of terrains is crucial for operational safety. The study proposes a real-time terrain classification method for delivery robots using MPU6050 IMU sensors to lower costs. Kalman filters are applied as pre-processing to reduce sensor noise, which significantly improves signal quality. The Artificial Neural Network (ANN) model is trained separately for each speed level (0.1–0.5 m/s) to handle vibration variations. The results show models with Kalman filters consistently outperform the raw data with an accuracy above 90%. Although real-time accuracy on ambiguous terrain initially varies between 66%–90%, the implementation of validation three times in a row has significantly improved system performance. Transition testing proves that the system is able to identify terrain changes and adjust the speed smoothly, ensuring navigation stability and safety of the robot's load.