Reinforcement learning for object recognition and room classification in an indoor environment

This paper presents an integrated system for autonomous navigation and object prioritisation using a mobile robot in an indoor environment. The system combines deep learning for room classification (VGG16) and object detection (YOLOv8) with Proximal Policy Optimisation (PPO) reinforcement learning to enable the robot to efficiently locate a target object (yellow duck) while avoiding distractions (tennis ball, lemon, banana). The robot operates in a Gazebo simulation with ROS2 Humble, leveraging Python for implementation. The VGG16 model was trained on bag-file-derived images to classify rooms (kitchen/dining area), while YOLOv8 was fine-tuned on annotated datasets in RoboFlow. PPO was employed to overcome challenges faced with Q-learning, optimising the robot’s path-planning and decision-making. Experimental results demonstrate the system’s ability to prioritise the target object with high accuracy, showcasing its potential for applications in service robotics and smart environments