Fast object approximation for real-time 3D obstacle avoidance with biped robots

Abstract

In order to achieve fully autonomous humanoid navigation, environment perception must be both fast enough for real-time planning in dynamic environments and robust against previously unknown scenarios. We present an open source, flexible and efficient vision system that represents dynamic environments using simple geometries. Based only on onboard sensing and 3D point cloud processing, it approximates objects using swept-sphere-volumes while the robot is moving. It does not rely on color or any previous models or information. We demonstrate the viability of our approach by testing it on our human-sized biped robot Lola, which is able to avoid moving obstacles in real-time while walking at a set speed of 0.4m/s and performing whole-body collision avoidance.