The benefit of muscle-actuated systems : internal mechanics, optimization and learning

Abstract

We are facing the challenge of an over-aging and overweight society. This leads to an increasing number of movement disorders and causes the loss of mobility and independence. To address this pressing issue, we need to develop new rehabilitation techniques and design innovative assistive devices. Achieving this goal requires a deeper understanding of the underlying mechanics that control muscle-actuated motion. However, despite extensive studies, the neural control of muscle-actuated motion remains poorly understood. While experiments are valuable and necessary tools to further our understanding, they are often limited by ethical and practical constraints. Therefore, simulating muscle-actuated motion has become increasingly important for testing hypotheses and bridge this knowledge gap. In silico, we can establish cause-effect relationships that are experimentally difficult or even impossible to measure. By changing morphological aspects of the underlying musculoskeletal structure or the neural control strategy itself, simulations are crucial in the quest for a deeper understanding of muscle-actuated motion. The insights gained from these simulations paves the way to develop new rehabilitation techniques, enhance pre-surgical planning, design better assistive devices and improve the performance of current robots. The primary objective of this dissertation is to study the intricate interplay between musculoskeletal dynamics, neural controller and the environment. To achieve this goal, a simulation framework has been developed as part of this thesis, enabling the modeling and control of muscle-actuated motion using both model-based and learning-based methods. By utilizing this framework, musculoskeletal models of the arm, head-neck complex and a simplified whole-body model are investigated in conjunction with various concepts of motor control. The main research questions of this thesis are therefore: 1. How does the neural control strategy select muscle activation patterns to generate the desired movement, and can we use this knowledge to design better assistive devices? 2. How does the musculoskeletal dynamics facilitate the neural control strategy in accomplishing this task of generating desired movements? To address these research questions, this thesis comprises a total of five journal and conference articles. More specifically, contributions I-III of this thesis focus on addressing the first research question which aims to understand how voluntary and reflexive movements can be predicted. First, we investigate various optimality principles using a musculoskeletal arm model to predict point-to-manifold reaching tasks. By using predictive simulations, we demonstrate how the arm would move towards a goal if, for example, our neural control strategy would minimize energy consumption. The main finding of this contribution shows that it is essential to include muscle dynamics and consider tasks with more openly defined targets to draw accurate conclusions about motor control. Through our analysis, we show that a combination of mechanical work, jerk and neuronal stimulation effort best predicts point-reaching when compared to human experiments. Second, we propose a novel method to optimize the design of exoskeleton power units taking into account the load cycle of predicted human movements. To achieve this goal, we employ a forward dynamic simulation of a generic musculoskeletal arm model, which is first scaled to represent different individuals. Next, we predict individual human motions and employ the predicted human torques to scale the electrical power units employing a novel scalability model. By considering the individual user needs and task demands, our approach achieves a lighter and more efficient design. In conclusion, our framework demonstrates the potential to improve the design of individual assistive devices. The third contribution focuses on predicting reflexive movements in response to sudden perturbations of the head-neck complex. To achieve this, we conducted experiments in which volunteers were placed on a table while supporting their heads with a trapdoor. This trapdoor was then suddenly released leading to a downward movement of the head until the reflexive reaction of the muscles stops the head from falling. We analyzed the results of these experiments, presenting characteristic parameters and highlighting differences between separate age and gender groups. Using this data, we also set up benchmark validations for a musculoskeletal head-neck model, including reflex control strategies. Our main findings are that there are large individual differences in reflexive responses between participants and that the perturbation direction significantly affects the reflexive response. Furthermore, we show that this data can be used as a benchmark test to validate musculoskeletal models and different muscle control strategies. While the first three contributions focus on the research question (1), contributions IV-V focus on (2) whether and how the musculoskeletal dynamics facilitate the learning and control task of various movements. We utilize a recently introduced information-theoretic approach called control effort to quantify the minimally required information to perform specific movements. By applying this concept, we can for example quantify how much biological muscles reduce the neuronal information load compared to technical DC-motors. We present a novel optimization algorithm to find this control effort and apply it to point-reaching and walking tasks. The main finding of this contribution is that the musculoskeletal dynamics reduce the control effort required for these movements compared to torque-driven systems. Finally, we hypothesize that the highly nonlinear muscle dynamics not only facilitate the control task but also provide inherent stability that is beneficial for learning from scratch. To test this, we employed various learning strategies for multiple anthropomorphic tasks, including point-reaching, ball-hitting, hopping, and squatting. The results of this investigation demonstrate that using muscle-like actuators improves the data-efficiency of the learning tasks. Additionally, including the muscle dynamics improves the robustness towards hyperparameters and allows for a better generalization towards unknown and unlearned perturbations. In summary, this thesis enhances existing methods to control and learn muscle-actuated motion, quantifies the control effort needed to perform certain movements and demonstrates that the inherent stability of the muscle dynamics facilitates the learning task. The models, control strategies, and experimental data presented in this work aid researchers in science and industry to improve their predictions in various fields such as neuroscience, ergonomics, rehabilitation, passive safety systems, and robotics. This allows us to reverse-engineer how we as humans control movement, uncovering the complex relationship between musculoskeletal dynamics and neural controller.