One of the special characteristics that distinguishes humanity from many other animals is our bipedal, upright gait. Our method of locomotion has evolved over millions of years, as have those methods other animals use, and many of these are now being studied very carefully for use in robots. Researchers at Cornell University however decided to approach walking a little differently by applying evolutionary algorithms on virtual robots to develop their gait or gallop.
Evolutionary algorithms work by mimicking natural evolution with mutations being introduced and the most successful systems being allowed to reproduce. In this case, being the most successful meant being the fastest. After 1000 generations, the robots proved to be unique looking with an interesting variety of mechanical solutions. Possibly with more generations the designs will be more similar to what we can find around us, as Nature has had the benefit of many times more generations.
The paper by the researchers concerning these robots can be freely downloaded: Unshackling Evolution: Evolving Soft Robots With Multiple Materials and a Powerful Generative Encoding (PDF).