Learning-based Biomimetic Strategies for Developing Control Schemes for Lower Extremity Rehabilitation Robotic Devices
by Sharmita Dey
Date of Examination:2023-02-07
Date of issue:2023-06-16
Advisor:Prof. Dr Arndt Schilling
Referee:Prof. Dr Arndt Schilling
Referee:Prof. Dr. Ulrich Sax
Referee:Prof. Dr. Dr. Dario Farina
Referee:Prof. Dr. Stephan Huckemann
Referee:Prof. Dr. Ramin Yahyapour
Referee:Dr. Sabri Boughorbel
Files in this item
This file will be freely accessible after 2024-02-06.
EnglishLower limb disabilities caused by factors such as amputation, neuromuscular impairment, and traumatic injury can significantly impact mobility and quality of life. To restore impaired functionality, individuals may use assistive devices such as prostheses, orthoses, or exoskeletons. However, most of the commercially available lower-limb prosthetic devices are passive and do not provide adequate energy or range of motion compared to natural limbs. This leads to compensatory movements, increased moments on the intact side, and fatigue, especially when performing high-energy tasks like stair ambulation, walking up inclines, and high-speed walking. With advancements in technology, motorized prosthetic devices have gained attention because their embedded motors can be actively actuated to produce the desired motion. Such motorized prostheses have the potential to support close-to-human-level movements by providing sufficient energy or a joint motion profile similar to that of a biological limb. Generating human-level movement with a motorized prosthesis is not a standalone phenomenon, but must be in tandem with the motion of the entire body. For this, a control strategy is required that incorporates how the natural limb synergy in humans functions and evolves across varying motion conditions. The representation of natural synergy is challenging because human gait is a complex coordination of multiple states induced by the neuromusculoskeletal system, and the explicit enumeration of all these states is practically intractable. Therefore, this thesis investigated the possibility of approximating the synergy exhibited during human gait as a digital twin by learning it from multiple human demonstrations. The synergy-learned digital twin model can thereafter be used to predict gait commands for operating a prosthetic limb such that the prosthetic limb functions similarly to a human limb. To this end, first, the feasibility of a synergy-learned gait-predictive model for generating accurate limb motion trajectories was assessed. Second, given that the reference joint motion to learn from is absent in disabled individuals, different training paradigms were explored to extrapolate the synergy-learned gait-predictive model to predict the missing limb motion for these individuals. Third, strategies for the adaptation of these models to temporal changes in gait patterns were explored by creating a convex combination of generic inter-individual models and adaptive individual-specific models using incremental learning. Fourth, the synergy-aware gait-predictive models were extended to jointly learn multiple locomotion tasks, such as different inclines and stair ambulation. To achieve this, different model architectures with fully-shared and partially-shared parameter spaces were explored. Furthermore, to enable the gait-predictive models to sequentially adapt to newly encountered gait tasks, a progressive model was created by incrementally adding new task-specific layers. To allow adaptation to deviations in the predictions and prevent the forgetting of previously learned gait tasks, a novel error-informed replay buffer was proposed. These approaches were evaluated on different publicly available datasets and our own laboratory-acquired motion-capture datasets. Additionally, to develop robust gait-predictive models for deployment, the models were trained using wearable sensor data from various motion conditions and the transitions between them to reflect natural gait patterns with turns, varying inclines, speeds, and mixed cadences. Furthermore, the trained models were prepared for real-time prediction by compressing them to reduce storage and time complexity. Finally, the gait-predictive models were deployed in real-time on an ankle exoskeleton, and able-bodied human subjects were able to walk comfortably using the synergy-learned model-predicted joint motion profiles. The adaptive gait-predictive models developed in this thesis offer new potential for the control of biomimetic prosthetic limbs and open promising directions in the field of assistive and rehabilitation robotics.
Keywords: Rehabilitation Robotics; Assistive Robotics; Exoskeleton; Prostheses; Machine Learning; Control