State Estimation for Legged Robots - Kinematics, Inertial Sensing, and Computer Vision

Abstract

The primary aim of this thesis is to endow legged robots with a reliable sense of ego-motion. Just like humans or other legged beings, many legged robots require estimates of their posture and velocity in order to keep balance and move through the environment. The estimates need to exhibit both sufficiently high bandwidth and accuracy in order to allow for a controlled execution of these tasks. Furthermore, due to this dependency, failures of the state estimation may quickly lead to damaging of the robot or its surroundings, which emphasizes the importance of the reliability of the employed estimation algorithms. The associated research question may be formulated as finding an appropriate combination of sensor modalities and state estimation algorithms such that the ego-motion can reliably and accurately be estimated with financially and computationally reasonable costs. Furthermore, the state estimation should not limit the capabilities of the robot and thus the use of restrictive assumptions such as a horizontal terrain, specific gait patterns, or the availability of external sensing is undesirable.