Vision-based foothold contact reasoning using curved surface patches

Abstract

Reasoning about contacts between a legged robot's foot and the ground is a critical aspect of locomotion in natural terrains. This interaction becomes even more critical when the robot must move on rough surfaces. This paper presents a new visual contact analysis, based on curved patches that model local contact surfaces both on the sole of the robot's foot and in the terrain. The focus is on rigid, flat feet that represent the majority of the designs for current humanoids, but we also show how the introduced framework could be extended to other foot profiles, such as spherical feet. The footholds are localized visually in the environment's point cloud through a fast patch fitting process and a contact analysis between patches on the sole of the foot and in the surrounding environment. These patches aim to compose a spatial patch map for contact reasoning. We experimentally validate the introduced visionbased framework, using range data for rough terrain stepping demonstrations on the COMAN and WALK-MAN humanoids.