Locomotion and Pose Estimation in Compliant, Torque-Controlled Hexapedal Robots

Kurzfassung

Several scenarios, such as disaster response or terrestrial and extra-terrestrial exploration, comprise environments that are dangerous or even inaccessible for humans. In those cases, autonomous robots pose a promising alternative to render such endeavours possible. While most of today's robotic explorers are wheeled or tracked vehicles, legged systems gained increased attention in recent years. With their unique combination of omnidirectional mobility and intrinsic manipulation capabilities, they are envisioned to serve as the rough terrain specialists in scouting or sample and return missions. Especially, small to mid-size hexapods are of great interest for those scenarios. Providing static stability across a wide range of walking speeds, they offer an attractive trade-off between versatility and complexity. Another important advantage is their redundancy, allowing them to tolerate the loss of single legs. However, due to their small size, the computational on-board resources are limited. Thus, the use of smart and efficient algorithms is of utmost importance in order to enable autonomous operation within a priori unknown rough environments. Working towards such autonomous robotic scouts, this thesis contributes with the development, implementation, and test of a self-contained walking layer as well as a 6 degrees of freedom (DOF) leg odometry for compliant, torque-controlled, hexapedal robots. Herein, the important property of all presented algorithms is the sole use of proprioceptive measurements provided by the legs, i.\,e.\ joint angles and joint torques. Especially the joint torque sensors improve the walking process by enabling the use of sensitive compliance controllers and distributed collision detection. Comprising a set of algorithms, the walking layer organises and structures the walking process in order to generate robust, adaptive, and leg loss tolerant locomotion in uneven terrain. Furthermore, it encapsulates the walking process, and thus hides its complexity from higher-level algorithms such as navigation. Its three main functional components are a flexible, biologically-inspired gait coordination algorithm, single leg reflexes, and active joint compliance control. Thereof, the gait coordination algorithm realises temporal adaptation of the step sequence while reflexes adjust the leg trajectories to the local terrain. The joint compliance control reduces internal forces and allows for situation dependent stiffness adjustments. An algorithmic extensions to the basic gait coordination enables the immediate adaptation to leg loss. In combination with stiffness and pose adjustments, this allows the hexapod to retain stable locomotion on five legs. In order to account for the emergent gait, the leg odometry algorithm employs an optimisation approach to obtain a kinematics-based pose estimate from joint angle measurements. Fusing the resulting pitch and roll angle estimates with joint-torque-measurement-based attitude data, reduces the associated drift, and thus stabilises the overall pose estimate. Various simulations and experiments with the six-legged, torque-controlled DLR Crawler demonstrate the effectiveness of the proposed walking layer as well as the 6-DOF leg odometry.