ABSTRACT
Theo Jansen linkage is an appealing mechanism to implement a bio-inspired motion for a legged robot. The oval orbit generated by a Theo Jansen linkage, possessing a transverse axis longer than a vertical axis, achieves an energy efficient walking as comparing to the circular orbit generated by a four-bar linkage. However, the ensemble of its links can produce different patterns of orbits other than oval orbits, some of which are not qualified to be the foot trajectories. It is vital to give a guideline, to which one can refer, to ensure the design of a Theo Jansen leg always possessing its eligibility. In contrast to the conventional approach of tabulating feasible design data, the machine learning technique called SVM (Support Vector Machine) serving as a classifier to distinguish desired trajectories from undesired ones is employed in this paper. Because it is not a linearly separable problem, the kernel method is applied to embed the input space into a higher-dimensional space, referred to as feature space or kernel space, in which the patterns of orbits can be linearly separable. Based upon SVM to delimit the eligible designs, one can seek the improvement of a Theo Jansen linkage by resizing its links without rendering an ineligible design.
