Machine Learning-Enabled Rover For Autonomous Navigation In Rough Terrain

A hybrid perception-control framework for navigating rough terrains by remote YOLOv8n-based terrain classification combined with on-board safety-based control of rover is proposed. The system uses ESP32-WROOM-32 module for controlling motors, ultrasonic sensors for detecting obstacles, infrared sensors for detecting cliff edges, and watchdog communication, whereas ESP32-CAM is responsible for sending terrain images to a laptop where the inference algorithm works. Four-class terrain data consisting of 800 images of grass, sand, gravel, and soil is employed for fine-tuning YOLOv8n using stratified dataset. Terrain predictions control movements depending on traversability coefficients, while safety controls take precedence over remote controls when an obstacle or communication link is detected. The evaluation over 60 trials of outdoor navigation has resulted in terrain classification with an accuracy of 78.3% and successful navigation with an 80.0% success rate. The perception-to-action latency was 170±23 ms, fail-safe time was 1.02±0.04 s, and emergency braking was 310±45 ms. The power consumption of the rover was 4.22±0.35 W, and the estimated operational time was around 4.5 h. Such values characterize the performance of the limited resources rover that implements the separation of complex computation-based inference from motion control in embedded systems.