The University of Edinburgh -
Division of Informatics
Forrest Hill & 80 South Bridge

MSc Thesis #94131

Title:Control of Walking by Central Pattern Generators
Date: 1994
Abstract:This work builds on that of [Taga et al. 91] in looking at the possibility of controlling robotic walking using neural networks. Taga's simulator has been reimplemented and his claim that the walking behaviour is "stably realized under variable conditions" has been investigated. Specifically, the ability of the legs to walk on different slopes and surfaces has been examined. Initial results showed that the legs could only walk on smooth ground and gentle slopes, but it became apparent that the poor modelling of the ground might be partially responsible for this. As a result the ground has been completely remodelled using coulombic friction, and the legs have been retested. It was found that the coefficient of friction of the ground could be varied a great deal without the legs falling over, although it cannot walk as fast on more slipper surfaces (as you would expect); overall performance is better, and, together with the added realism, this is a very pleasing result. Two further changes have also been made. Firstly the dynamics algorithm was altered: initially the number of variables was reduced in the algorithm to increase its accuracy and efficiency, but then the algorithm itself was changed to use an adaptation of the Composite Rigid Body technique described in [Walker and Orin 82, Featherstone 84]. This increased the ease with which the model could be redesigned in the future, both by making the algorithm more explicit and by allowing the model to be three dimensional. Secondly the model itself was changed to include a rotating hip joint which it was hoped would allow the feet to clear the ground more easily when swinging forwards. Unfortunately, it was not possible to design a suitable control mechanism for this in the time available.

[Search These Pages] [DAI Home Page] [Comment]