DEEP REINFORCEMENT LEARNING BASED AUTONOMOUS NAVIGATION ROBOT FOR PATIENT FOLLOWER DRIP-HOLDER

This research principally focused on design an autonomously navigation mobile robot system based on deep reinforcement learning (DRL/DL) called autonomous patient follower robot (APFR). APFR will be used in hospital and rehabilitation centers to help patients as well as hospital staff by holding patient’s drip stand and move around by avoid obstacles with in static and dynamic hospital environment with appropriate slack distance. To have the autonomy of the APFR, a four-wheeled navigation robot with four DC motors for each wheel & NVIDIA artificial intelligence (AI) board called NVIDIA Jetson Nano 2 GB developer kit (NJNDK) is utilized as the heart and brain of the system. With the help of the DRL computing algorithm inside the developer kit, the robot is capable of avoiding obstacles, distance estimation, and path planning tasks. The algorithm that we have employed is called stochastic gradient descent (SGD) algorithm that is a part of gradient descent algorithm. SGD is employed because of the algorithm takes less computing time compare to other gradient descent algorithms. The ANN architecture with the SGD algorithm gives the robot able to reroute its motion direction to the patient and avoid obstacles. To train the APFR, data (image) is collected within the model environment and trained using ResNet18 & AlexNet Pre-trained CNN architectures with SGD/DRL-algorithm. By utilizing the SGD-algorithm for APFR allows the robot to be flexible, reliable, nimble, and faster to recognize and learn its environment (dynamic or static).