Daniel Flower Dot Com Banner


In this project a virtual city was implemented in order to simulate vision guided navigation of autonomous agents. The agents were given no information about the layout of the world or which objects existed there, and instead rendered the scene from their points of view and decided how to move based on this information, much as a robot would do situated in the real world.

Specifically, the pixel values of the rendered scene were used as inputs into a neural network, with the output of the neural network giving information on how the agent should be moved. The agents were separately trained to avoid obstacles, follow the road, obey traffic lights, and path find. Training consisted of a human trainer controlling the agents as in a video game, with a log of the rendered scene and the action the human controller was taking at each time step. Separately, the neural network was trained using the data from these log files.

The results of this project were mixed, with very good performance in obstacle avoidance and path following, but poor performance in obeying traffic lights. The path finding, which involves both obstacle avoidance and travelling towards a goal destination, showed much promise. It is probable that with further training the neural network could learn to do this better, and it was speculated that in autonomous agents situated in the real world, neural network-based visual guidance might be suitable for moving the agent to a goal destination, with this destination being decided by other higher-level methods. This is because the neural network is able to learn in environments where lots of noisy data is present, and there are many variations in the scene, because with enough training the neural network is able to generalise out the important parts of a scene in a way that would be very difficult with hand coded algorithms.