Animals and humans are constantly faced with a highly dimensional stream of incoming sensory information. At the same time, they have to command their highly complex and multidimensional bodies. Yet, they seamlessly cope with this situation and successfully perform various tasks. For autonomous robots, this poses a challenge: robots performing in the real world are often faced with the curse of dimensionality. In other words, the size of the sensory as well as motor spaces becomes too large for the robot to efficiently cope with them in real time. In this paper, we demonstrate how the curse of dimensionality can be tamed by exploiting the robot’s morphology and interaction with the environment, or the robot’s embodiment (see e.g., [1]). We present three case studies with underactuated quadrupedal robots. In the first case study, we look at terrain detection. While running on different surfaces, the robot generates structured multimodal sensory information that can be used to detect different terrain types. In the second case study, we shift our attention to the motor space: the robot is learning different gaits. The online learning procedure capitalizes on the fact that the robot is underactuated and on a “soft“ control policy. In the third case study, we move one level higher and demonstrate how - given an appropriate gait - a speed adaptation task can be greatly simplified and learned online.
展开▼
