I have done few changes since my last (MSLS) project. Now I used much simpler neural network, which proved to be sufficient. Genetic algorithm has been improved by using elitism. Graphics was completely changed to my own small "openGL engine".
I have run 15 evolutionary runs and tested rover in many different terrains. The neural network controller is much more robust than its predecessor's.
This video will be played on some plasma screens across our university so if you have any ideas how to improve, please let me know.
In the research that is shown in this video, the neural controller learned to avoid obstacles in the simplest way to maximise the fitness given by specific fitness function. There are several scientific papers that we published (check my website to find them). The rover in this initial research setup learned to avoid to turn to one side only, which was not the case after we introduced more complex modalities such as the active vision system passed to the neural net.
music is clubbed to death
mpeniak 1 year ago