Automatic Generation Of Legged Robotic Controllers For Planetary Exploration
As NASA has moved towards unmanned space exploration, autonomous vehicles that can navigate the surfaces of planets like Mars have played an increasingly prominent and essential role. An important goal for future exploration is to expand the scope of areas that are accessible to such unmanned exploration. One potential path towards such expanded access is to replace wheeled robots, which can become stuck or lose grip in difficult terrain, with robots that walk on legs. However, the problem with legged robots is that they can be expensive to design. Our SRI project aims accordingly to open space exploration to legged robots (in particular
quadrupeds) by significantly diminishing the cost of developing controllers for such robots. In particular, the idea is to apply a machine learning technique called an evolutionary algorithm to generate a controller for any particular legged robot morphology automatically so that the appropriate controller does not need to be designed or calibrated for the particular body by human engineers. While the obstacles to such automatically-generated controllers are significant, our group is developing specialized representations of ambulatory control mechanisms based on artificial neural networks that simplify the problem enough to make it feasible. This talk will detail the technology behind this initiative and progress so far towards its achievement.