Using it's onboard sensors, and going through a bootstrapping phase where it moves each of it's joints and checks the results, it builds a virtual model of itself. Then, through a stage of using random movements, it teaches itself a means of locomotion. As the robot teaches itself, it can relearn how to walk should it be damaged, for instance missing a joint.
Further information is available on the Cornell CCSL Robotics Self Modelling web site.