## AbstractBiological systems exhibit an impressive ability to interact with their environment. They display high versatility in their movements, an ability to learn fast, and a remarkable robustness to noise and external perturbations. On the other hand even modern robots often look clumsy, have difficulty to learn successfully in noisy, dynamic environments, and are prone to perturbations. Therefore, robot designers seek for inspiration from nature by identifying successful strategies of biological systems and applying them to robots. With this dissertation I want to contribute to further close the performance gap between artificial robots and their biological role models. My approach is to employ experimental data, that have been collected over a wide range of biological systems and over a variety of different tasks. I identify the underlying general strategies, formalize them by the use of rigorous mathematics, to finally apply them to robots. My dissertation is divided into two main parts, dealing with two different biologically inspired approaches. The first one, kinematic synergies, describes the phenomenon in biological systems that a number of degrees of freedom (e.g., muscles, joints) are coordinated in a fixed way in order to fulfill a single task. This allows the controller, which operates the synergy, to work in lower dimensional space. Hence, the remaining control and/or learning task is much simpler. I demonstrate how kinematic synergies can be formulated mathematically and how they can be applied to balance control of a humanoid robot. Remarkably, such synergies, in conjunction with simple linear controllers, enable a humanoid robot to balance online against all kinds of unknown, dynamic perturbations with very little computational costs. The second biologically inspired strategy, which I describe, is known as morphological computation. It embraces the observation that the physical body (i.e., the morphology) is not simply a device to carry the brain around, but rather that it is highly involved in computational tasks. There exit already a number of robots, which successfully implement this concept. Nevertheless, a theoretical basis for understanding the capabilities and limitations of morphological computation has been missing so far. I present different theoretical models for morphological computation, where a precise mathematical characterization of the potential computational contribution of a complex physical body is feasible. Based on these models I propose morphological computation setups, which consist of the physical body itself with static readouts and static feedbacks. These simple structures are able to emulate a surprisingly rich class of nonlinear computations (even ones with persistent memory), which map continuous input streams to continuous output streams. Remarkably, in the case of a complex, compliant body it is sufficient to add linear outputs and linear feedbacks in order to emulate nonlinear, dynamic computations. This points to an interesting property of morphological computation: it facilitates learning. By outsourcing parts of the computation to the physical body, the complex problem of learning to control a robot may be reduced to the much simpler task of finding linear output weights. This suggests that complexity and nonlinearity, typically unwanted properties of robots, are desired features in order to provide a computationally powerful physical body.
[Edit] |