On accuracy, robustness and reliability of laser-based localization

Abstract

Legged robots are expected to demonstrate autonomous skills in situations which are not suitable for wheeled platforms, such as cluttered disaster areas and outdoor trails. State estimation strategies for legged robots are typically based on inertial sensing and leg kinematics, and are affected by continuous drift. In a process called localization, perception sensors such as cameras or laser scanners can be used to compensate for this drift via scene registration. The objective of this dissertation is to develop laser-based localization techniques which are suitable for the safe and continuous operation of wheeled and legged robots in real-world environments.We explore approaches fusing inertial, kinematics, stereo vision and laser signal sources in various combinations, in order to enable state estimation in the presence of challenges such as limited sensor field-of-view, varying lighting conditions, occlusions and non-uniform spatial overlap during motion. A number of novel contributions are presented, including: a method for laser-based localization which is robust to large variations in spatial overlap and initial alignment error, a state estimation system which fuses multiple sensor sources in a manner which is suitable for a quadruped’s feedback controller, a module which can predict localization failure. The final contribution, explores the applicability of deep learning techniques to laser odometry. We carry out extensive experimental evaluations running real-time in closed loop control on the NASA Valkyrie humanoid robot and the IIT quadruped HyQ, while exploring a variety of test environments. We run further experiments using a dataset collected with the Boston Dynamics humanoid Atlas during the DARPA Robotics Challenge finals, as well as on two mobile robots, using ours and publicly available datasets.