• convolutional neural network
• foothold selection
• footstep planning
• legged locomotion
• quadruped robots

Legged locomotion over rough and unstructured terrain is a highly dynamic, nonlinear, and contact-rich task. This challenging scenario is typically tackled relying on the planning of the foothold position to maintain grip on the terrain and thus guarantee the stability of the robot. Recently, several types of feet are starting to be used on quadruped robots, enabling high-grip contacts for different types of surfaces and terrains. However, little work has been done on controllers and planners that are able to exploit their different capabilities, and most of the control approaches works on the point-contact foot hypothesis. In this work, I present a foothold optimization system that can search for the optimal contact points for feet with different shapes, using a polynomial approximation. The proposed system uses a fully convolutional network trained on simulated data to predict a cost index for any candidate foothold. The optimization system is experimentally validated for point-shaped feet and soft feet on the ANYmal-C robot.