Evolving virtual creatures revisited

Thirteen years have passed since Karl Sims published his work on evolving virtual creatures. Since then, several novel approaches to neural network evolution and genetic algorithms have been proposed. The aim of our work is to apply recent results in these areas to the virtual creatures proposed by Karl Sims, leading to creatures capable of solving more complex tasks. This paper presents our success in reaching the first milestone - a new and complete implementation of the original virtual creatures. All morphological and control properties of the original creatures were implemented. Laws of physics are simulated using ODE library. Distributed computation is used for CPU-intensive tasks, such as fitness evaluation. Experiments have shown that our system is capable of evolving both morphology and control of the creatures resulting in a variety of non-trivial swimming and walking strategies. Categories and Subject Descriptors I.2.9 [ARTIFICIAL INTELLIGENCE]: Robotics; I.2.8 [ARTIFICIA...

• Type: Conference
• Year: 2007
• Where: GECCO
• Authors: Peter Krcah
• Keywords: GECCO 2007, Karl Sims, Original Virtual Creatures, Virtual Creatures