Rough terrain mapping and classification for foothold selection in a walking robot

Although legged locomotion over a moderately rugged terrain can be accomplished by employing simple reactions to the ground contact information, a more effective approach, which allows predictively avoiding obstacles, requires a model of the environment and a control algorithm that takes this model into account when planning footsteps and leg movements. This article addresses the issues of terrain perception and modeling and foothold selection in a walking robot. An integrated system is presented that allows a legged robot to traverse previously unseen, uneven terrain using only onboard perception, provided that a reasonable general path is known. An efficient method for real‐time building of a local elevation map from sparse two‐dimensional (2D) range measurements of a miniature 2D laser scanner is described. The terrain mapping module supports a foothold selection algorithm, which employs unsupervised learning to create an adaptive decision surface. The robot can learn from realistic simulations; therefore no a priori expert‐given rules or parameters are used. The usefulness of our approach is demonstrated in experiments with the six‐legged robot Messor. We discuss the lessons learned in field tests and the modifications to our system that turned out to be essential for successful operation under real‐world conditions. © 2011 Wiley Periodicals, Inc.