der-tim.de

Every year, the new cognitive science students in Osnabrück design a fancy logo and print it on all kinds of cloth you can imagine (ok, basically on shirts, but who cares? ;-) ). When I started my undergraduate studies three years ago, we chose the logo you can see below – the evil robot. (Don’t worry, in the read more-part I won’t only talk about t-shirts. Its more about robots ;-) )

CogSci Logo 2006

Right now, I am doing my neurocybernetics homework. Using a nice little simulator which is beeing developed at our university, I try to create different kinds of neural networks to control a humanoid robot. In the end (at least I hope so), the robot should do something that looks like walking.

Yes, I know, at the moment it looks more like a robot on drugs in space… But I hope the guy will walk more naturally in some days (i.e. tomorrow). Ans I hope to have the time to also create a system that really walks. For humans, walking is a quite simple activity – at least in most cases. We can step forward automatically, and, normally, do not stumble or fall down. But modelling this simple behaviour is not that easy. Taking into account the influence of physics, such as friction of the ground, inertia and gravity, and making a robot react adequately actually is a hard task. In the simulator we use, the robot is entirely controlled by neural networks. They get input from some sensor neurons (measuring angle and acceleration of some limbs and joints) and can output to neurons to set a desired angle or torque of the limbs and joints. So it is impossible to directly write a control program or mathematical function to create the movement. Instead, the function needs to be encoded in a neural net. One advamtage of this approach is that we can use evolutionary algorithms to create better control networks. Designing a net by hand which controls the 48 output neurons and finding the right parameters to let the robot walk human-like would maybe take some years. (That is at least my naive impression after having worked with this stuff for some months now.) But if you can design a suitable, but nevertheless simple initial network and specify a criterion for networks being good or bad, you can use evolution to do the job. For this, a population of several networks is created. These networks ore slight mutations of your original net, varying for example in the number or strength of synapses, or number of hidden neurons of the net. Then, all networks of the population are tested and rated by a fitness function. This function could, for example, reward nets that make a robot walk and punish the ones that let him fall down. After all networks have been tested, the best individuals are selected and form the parents of a new generation, which again may contain mutations of these networks. After some time – depending on the difficulty of the task, the initial net and the chosen parameters – evolution may create networks which let the robot walk human-like.

However, there is another point I wanted to make (if anyone who reads this is interested in some of the concepts mentioned above, just drop me a line in the comments – then I’d write something more about the specific topic). When I read thepost at Neurotopia this afternoon, I was remembered that we do not build the robots. Unfortunately. It would have been nice to see how such a little toy would be controlled by the programs we are developing. Some more or less little neural networks controlling a one-ton 18-foot tall robot. Maybe we should ask the guy to collaborate with our department. That would be definitely more fun than programming humanoid robots that are as big as a small dog – or not even real ¹ . Anyone out there who has a spare giant robot to donate to some curious students? Maybe Owens gives away his first prototype when having finished the new one. Then, does anyone know how much a one-ton packet from Canada to Germany via airmail costs?