Abstract: Foothold identification is a key ability for legged robots that allows generating terrain adaptive behaviors (e.g., gait and control parameters) and thereby improving mobility in complex environment. To this end, this paper addresses the issue of foothold characterization and identification over rugged terrain, from the terrain geometry point of view. For a terrain region that might be a potential foothold of a robotic leg, the characteristic features are extracted as two first-order partial derivatives and two curvature parameters of a quadric regression surface at this location. These features are able to give an intuitive and, more importantly, accurate characterization towards the specific geometry of the ground location. On this basis, a supervised learning technique, Support Vector Machine (SVM), is employed, seeking to learn a foothold identification policy from human expert demonstration. As a result, an SVM classifier is learnt using the extracted features and human-demonstrated labels, which is able to identify whether or not a certain ground location is suited as a safe foot support for a robotic leg. It is shown that over 90% identification rate can be achieved with the proposed approach. Finally, preliminary experiment is implemented with a six-legged robot to demonstrate the effectiveness of the proposed approach.

Key words: six-legged robot, foothold characterization, foothold identification, geometric characteristics, support vector machine