Autonomous robots, such as automatic vacuum cleaners, toy robot dogs, and autonomous vehicles for the military, are rapidly becoming a part of everyday life. As a result the need for effective algorithms to control these agents is becoming increasingly important. Conventional path finding techniques rely on a representation of the world that can be analysed mathematically to find the best path. However, when an agent is placed into the real world in a place it has not seen before, conventional techniques frequently fail and a fundamentally different approach to path finding is required. The agent must rely on its senses, such as the input from a mounted camera, using this information to get around. We are especially interested in algorithms for use in highly interactive virtual environments such as computer games. In this paper we devise and analyse a technique which enables autonomous agents to navigate their way around a virtual city by using only what they see from their point of view. Since the scenes are computer generated we can use for the player's view and the agent's view representations with different visual complexity and hence improve the efficiency and effectiveness of the neural network. We show that by using neural networks agents can learn how to avoid obstacles, to follow the road, and we demonstrate that this method might even be useful for integration in path finding algorithm.