Moving and adapting with a learning exoskeleton
Author(s)Siu, Ho Chit
Massachusetts Institute of Technology. Department of Aeronautics and Astronautics.
Leia A. Stirling.
MetadataShow full item record
The operation of a powered exoskeleton is a type of human-robot interaction with extremely tight human-robot coupling. As exoskeletons become increasingly intelligent, it is increasingly appropriate to think of them not simply as tools, but rather as semi-autonomous teammates. This thesis explores the implementation, operation, and consequences of intelligent exoskeletons - teammates that move and adapt to the human to which they are physically coupled. Exoskeletons have potential applications in several domains, including strength augmentation, injury reduction, and rehabilitation. Appropriately mapping human intent to exoskeleton action is crucial. Generating this mapping can be difficult, as operator movements are constrained by the exoskeletons they are trying to control. This problem is particularly significant with upper-body exoskeletons, where high degrees of freedom allow for much less predictable motion than in the lower body. Surface electromyography (sEMG) - reading electrical signals from muscles - is one way to estimate human intent. sEMG contains anticipatory information that precedes the associated limb movement, allowing for better human-exoskeleton coordination than reactive control methods. However, sEMG is very sensitive to individual physiologies and sensor placement. We use machine learning from demonstration (LfD) to create personalized, robust sEMG mappings for exoskeleton control. We demonstrate classification of transient dynamic grasping gestures with data where sEMG sensors on the forearm have been shifted from a nominal configuration. Next, sEMG-based gesture recognition is applied to exoskeleton control, where sEMG mappings are learned as the exoskeleton is controlled with a pressure-based inputs. Finally, we analyze the human-exoskeleton team performance, fluency, and adaptation using a pressure-based controller, a static sEMG mapping, and a dynamic sEMG mapping. We show that LfD allows us to use anticipatory signaling to reduce human-exoskeleton interaction pressure. Subjects were able to adapt to all three controllers, but team performance and fluency were affected by the controller type and order of exposure. These results have implications for future exoskeleton controller design, and for exoskeleton operator training. They also open up new avenues of research in relation to adaptation to exoskeletons, intent classification algorithms, and the application of metrics from the human-robot interaction literature to the field of human exoskeleton research.
Thesis: Ph. D., Massachusetts Institute of Technology, Department of Aeronautics and Astronautics, 2018.Cataloged from PDF version of thesis.Includes bibliographical references (pages 129-142).
DepartmentMassachusetts Institute of Technology. Department of Aeronautics and Astronautics.
Massachusetts Institute of Technology
Aeronautics and Astronautics.