Kapalı ortamlarda derin öğrenim tabanlı görsel navigasyon

Deep learning methods are used various areas in recent years. Quite efficient results are obtained in recent studies with combination of deep learning and reinforcement learning. Deep learning based reinforcement learning especially gives powerful solutions for complex robotic tasks like navigation. Mobile robots gains new skills with Deep Reinforcement Learning (DRL). In this thesis, we propose a deep reinforcement learning approach to complex navigation tasks in indoor environments. We chose an unmanned ground vehicle as an agent and performed visual navigation simulation with real-world camera images. Proximal Policy Optimization (PPO) is policy update method which we used. We investigated various kind of neural network models to find best function approximator such as Convolutional Neural Networks (CNN), Multi-layer Perceptron (MLP), Extreme Learning Machines (ELM), Residual Neural Networks (ResNet) and Neural Ordinary Differential Equations (ODEs). Up to the our knowledge, the use of ODEs with DRL in navigation applications has not proposed in the literature. Results show that ODE based DRL algorithm performs well and makes gain the capability of navigation to the agent in indoor environment.