Recent demographic trends such as the increase in elderly population has caused a severe shortage of caregivers. Assistive robots are a promising solution, that can improve quality of life for the elderly and reduce the burden on caregivers. The major requirements for such a caregiving robot includes safe human-robot interaction and the ability to manipulate a wide array of household items.
A major problem that arises with aging is loss of motor functions to perform dextrous tasks such as clothing. Assistance with dressing is an essential task that could potentially be performed by robots. However, robotic clothing assistance is considered an open problem with several challenges involved. The task of clothing assistance follows non-linear dynamics with large variabilities between different environmental conditions. There is also an emphasis on fast learning due to its application in a health-care setting wherein the robot needs to model the task using few observations in a data-efficient manner. Design of an efficient clothing assistance framework involves reliable cloth state estimation and data-efficient motor skills learning.
In this thesis, Bayesian nonparametric latent variable models are used to tackle two problems of robotic clothing assistance. Firstly, the problem of reliable cloth state estimation is addressed. Manifold Relevance Determination (MRD) is used to learn a shared latent space for observations from a noisy depth sensor and accurate motion capture system. This latent space is used to infer the accurate cloth state given only the noisy depth sensor readings in a real-world setting. In the second case, the problem of data-efficient motor skills learning for clothing assistance is addressed. Bayesian Gaussian Process Latent Variable Model (BGPLVM) is used to learn a low dimensional latent space that can encode the task-specific motor skills for clothing assistance. It is demonstrated that performing policy search reinforcement learning in this latent space outperforms learning in the high-dimensional joint configuration space of the robot. This framework is also presented as a user-friendly tool that can be used to impart novel motor skills to bulky humanoid robots.
The proposed frameworks provide a promising solution towards realizing robotic clothing assistance. The thesis is concluded with some theoretical considerations in modeling the task of clothing assistance and a roadmap to achieve practical implementation in real-world scenarios such as a household environment.