Early phases of humanoid vehicle ingress using depth cameras

Abstract

This paper presents work on integrating perception and motion-planning for a humanoid robot to ingress, or enter, a small utility vehicle as a precursor to driving, a stage in the recent DARPA Robotics Challenge (DRC). Using a Hubo 2+ robot platform and a pair of RGB-D cameras, we describe a set of approaches to and present results on the first four phases of ingress: (1) visually search for the vehicle's doorway as a target to walk toward, (2) plan and execute a collision-free approach to the doorway via walking using visual odometry, (3) make visually-guided fine positioning adjustments near the entry door during docking, and (4) step from the ground up to the floor of the vehicle. All recognition is done on 3-D point clouds derived from the depth cameras without appearance information. Some further ingress results from the DRC-Hubo making use of grasping are also shown.