| Abstract [eng] |
In this work, two different artificial neural networks are trained and applied. First, the theory behind artificial intelligence, machine learning and different supervised and reinforcement algorithms is reviewed. Additionally, the tools used in this work are presented. The first neural network is used for object detection. An algorithm called YOLOv5 (You Only Look Once) is used for this purpose. The searched object is a square piece of metal. Once fully trained, the model reached a mAP (mean average precision) value of 99%, a recall and precision values of 1 are also reached. Practically, the model works very well, correctly identifying and finding the object but only at a 5 frame per second rate using a video feed from a camera due to insufficient computing resources. The second neural network used in this work is responsible for the control of a mechanical device - a robot arm. The algorithm used for this task is called PPO, or Proximal Policy Optimization. A virtual environment is created in Unity Software, where the training of the model happens using a plugin called ML-Agents. Training is not as successful due to the constraints of object controls presenting in Unity Software. The best result reached is a success rate of 3 out of 10 times of finding the object. While practically possible, transferring an artificial intelligence model from a virtual environment to a real robotic system proves difficult. The aim of this work is to train and apply an artificial neural network for image recognition and teaching a robot to find and touch an object in a virtual environment. |
