Synthesizing Physically-Realistic Environmental Models from Robot Exploration

In previous work [4] a framework was demonstrated that allows an autonomous robot to automatically synthesize physically-realistic models of its own body. Here it is demonstrated how the same approach can be applied to empower a robot to synthesize physically-realistic models of its surroundings. Robots which build numerical or other nonphysical models of their environments are limited in the kinds of predictions they can make about the repercussions of future actions. In this paper it is shown that a robot equipped with a self-made, physicallyrealistic model can extrapolate: a slow-moving robot consistently predicts the much faster top speed at which it can safely drive across a terrain.