Developing neural networks for the behavior control of autonomous robots can be a time-consuming task. This is especially the case for the new generation of complex robots with many sensors and motors – such as humanoid robots –, for which the networks with hundreds of neurons can become comparably large. Looking at the corresponding controller design workflow, a number of properties can be identified that slow down the development process: (1) The difficulty to create, handle and comprehend the large neuro-controllers, (2) the intricate debugging of neuro-controllers on the hardware, (3) delays caused by frequent time-consuming uploads of controllers to the hardware, (4) potential damaging of the robot and (5) the overall maintenance effort. This article proposes several measures to improve this workflow with respect to the mentioned problems. Some proposed improvements are realized by using sophisticated evolutionary robotics development software and suitable graphical network design tools. Such software, here in particular the Neurodynamics and Evolutionary Robotics Development Toolkit (NERD), significantly improves the network design process, specifically by allowing the development partially in simulation, by allowing a visual design of controllers with graphical network editors and by using suited neuro-evolution algorithms. Other improvements are based on proper neuro-modules that can be used to increase the usability of existing controllers. Bundled together, the proposed measures lead to a faster development of neuro-controllers. The proposed methods are demonstrated exemplarily with the Myon humanoid robot, but they can be applied also to other robots with similar properties and thus can help to improve the workflow for the neuro-controller design on such robot hardware.
Enhancing the Neuro-Controller Design Process for the Myon Humanoid Robot
2013-07-17
Paper
Elektronische Ressource
Englisch
Walking Motion Generation and Neuro-Fuzzy Control with Push Recovery for Humanoid Robot
BASE | 2017
|