Samenvatting
The recent development in robotics and human motion piqued interest in researching dynamic human-interactive devices. Specifically, many prototypical designs for exoskeletons, which provide physical assistance, have been proposed. These devices give support by applying strength in the direction the wearer moves to, in order to meet the motion requirements.However, current prototypes require the calibration of the actuators in order to provide person-specific support. In this thesis, we aim to set up a theoretical exoskeleton-assistance framework in which we learn the required support using Gaussian processes, rather than calibrating it ourselves. We introduce a novel method, called SEAM, to learn support in an on-line environment. Additionally, we explore the capabilities of an "all-in-one" support model, which considers the aggregation of multiple activities in one model. In this study, we found that, although conceptually applicable, the generative aspect of SEAM prevents a stable series of predictions, thus failing to learn a proper support model in an on-line environment. We also found that, even though a single-task support model performs significantly better than a multi-task support model in terms of predictive accuracy, the latter provides a much more compact and less redundant approach. Moreover, using a multi-task support model allows for extrapolation to other unseen tasks. Overall, this study provides useful insights in terms of modeling support provided by exoskeletons.
| Datum prijs | sep. 2015 |
|---|---|
| Originele taal | English |
| Prijsuitreikende instantie |
|
| Begeleider | Anna Harutyunyan (Advisor) & Ann Nowe (Promotor) |
Citeer dit
- Standard