Advancing Musculoskeletal Robot Design for Dynamic and Energy-Efficient Bipedal Locomotion

Achieving bipedal robot locomotion performance that approaches human performance is a challenging research topic in the field of humanoid robotics, requiring interdisciplinary expertise from various disciplines, including neuroscience and biomechanics. Despite the remarkable results demonstrated by current humanoid robots---they can walk, stand, turn, climb stairs, carry a load, push a cart---the versatility, stability, and energy efficiency of humans have not yet been achieved. However, with robots entering our lives, whether in the workplace, in clinics, or in normal household environments, such improvements are increasingly important. The current state of research in bipedal robot locomotion reveals that several groups have continuously demonstrated enhanced locomotion performance of the developed robots. But each of these groups has taken a unilateral approach and placed the focus on only one aspect, in order to achieve enhanced movement abilities;---for instance, the motion control and postural stability or the mechanical design. The neural and mechanical systems in human and animal locomotion, however, are strongly coupled and should therefore not be treated separately. Human-inspired musculoskeletal design of bipedal robots offers great potential for enhanced dynamic and energy-efficient locomotion but also imposes major challenges for motion planning and control. In this thesis, we first present a detailed review of the problems related to achieving enhanced dynamic and energy-efficient bipedal locomotion, from various important perspectives, and examine the essential properties of the human locomotory apparatus. Subsequently, existing insights and approaches from biomechanics, to understand the neuromechanical motion apparatus, and from robotics, to develop more human-like robots that can move in our environment, are discussed in detail. These thorough investigations of the interrelated essential design decisions are used to develop a novel design for a musculoskeletal bipedal robot, BioBiped1, such that, in the long term, it is capable of realizing dynamic hopping, running, and walking motions. The BioBiped1 robot features a highly compliant tendon-driven actuation system that mimics key functionalities of the human lower limb system. In experiments, BioBiped1's locomotor function for the envisioned gaits is validated globally. It is shown that the robot is able to rebound passively, store and release energy, and actively push off from the ground. The proof of concept of BioBiped1's locomotor function, however, marks only the starting point for our investigations, since this novel design concept opens up a number of questions regarding the required design complexity for the envisioned motions and the appropriate motion generation and control concept. For this purpose, a simulator specifically designed for the requirements of musculoskeletally actuated robotic systems, including sufficiently realistic ground reaction forces, is developed. It relies on object-oriented design and is based on a numerical solver, without model switching, to enable the analysis of impact peak forces and the simulation of flight phases. The developed library also contains the models of the actuated and passive mono- and biarticular elastic tendons and a penalty-based compliant contact model with nonlinear damping, to incorporate the collision, friction, and stiction forces occurring during ground contact. Using these components, the full multibody system (MBS) dynamics model is developed. To ensure a sufficiently similar behavior of the simulated and the real musculoskeletal robot, various measurements and parameter identifications for sub-models are performed. Finally, it is shown that the simulation model behaves similarly to the real robot platform. The intelligent combination of actuated and passive mono- and biarticular tendons, imitating important human muscle groups, offers tremendous potential for improved locomotion performance but also requires a sophisticated concept for motion control of the robot. Therefore, a further contribution of this thesis is the development of a centralized, nonlinear model-based method for motion generation and control that utilizes the derived detailed dynamics models of the implemented actuators. The concept is used to realize both computer-generated hopping and human jogging motions. Additionally, the problem of appropriate motor-gear unit selection prior to the robot's construction is tackled, using this method. The thesis concludes with a number of simulation studies in which several leg actuation designs are examined for their optimality with regard to systematically selected performance criteria. Furthermore, earlier paradoxical biomechanical findings about biarticular muscles in running are presented and, for the first time, investigated by detailed simulation of the motion dynamics. Exploring the Lombard paradox, a novel reduced and energy-efficient locomotion model without knee extensor has been simulated successfully. The models and methods developed within this thesis, as well as the insights gained, are already being employed to develop future prototypes. In particular, the optimal dimensioning and setting of the actuators, including all mono- and biarticular muscle-tendon units, are based on the derived design guidelines and are extensively validated by means of the simulation models and the motion control method. These developments are expected to significantly enhance progress in the field of bipedal robot design and, in the long term, to drive improvements in rehabilitation for humans through an understanding of the neuromechanics underlying human walking and the application of this knowledge to the design of prosthetics.