Anna Pawlicka

Programmer. Hiker. Cook. Always looking for interesting problems to solve.


About | Archive | Talks

A not very successful attempt at red ball fitness function.

02 Feb 2013 | AI, Aibo, artificial curiosity, reinforcement learning, robotics, Webots

I decided to try a very simple genetic algorithm, with my own methods to copy and mutate individuals, and a very basic fitness function that, in theory, should try to minimize the distance between Bazinga and the red ball.

Results:

Problem: Bazinga has no idea how to walk. It gets stuck when facing the wall. Its movements resemble jerks.

Possible solution 1: Input comes from 12 joints. Motor actions are applied to 12 joints as well. I will limit the number of outputs to e.g. 4.

Possible solution 2: Four legged animals perform synchronized movements. Should I implement that? They are also born with reflexes. Maybe it would be smarter to hardcode some basic reflexes?

Possible solution 3: The fitness function could measure the speed of getting from one place to another. This could theoretically evolve the most efficient way of moving. This may not resemble typical walking at first, but maybe eventually it would lead to it, e.g. lizards vs mice (their efficiency of locomotion is affected by the setting of their limbs).


Older · View Archive (25)

Bazinga the Retarded Dog

Since October last year I have been working on a robot simulation (Sony’s Aibo) trying to implement artificial curiosity and learning algorithms. Why? Final-year project. But also a very interesting and challenging problem to solve. I am working under supervision of Dr Chrisantha Fernando, at Queen Mary University of London.

Newer

How to add a new sensor to Aibo/NAO simulation in Webots

Most of the robot simulations can be modified by adding Nodes directly in Scene Tree in Webots itself. However, to add a new sensor to Aibo (and NAO), we need to modify proto file.