For robots that aim at providing physical assistance, a significant challenge is adapting to the different needs of people. A way to overcome this is to ask human experts to provide demonstrations of how to help a specific person; a setting known as Learning Assistance by Demonstration (LAD). In this thesis, we consider the application of robotic wheelchairs and investigate: how can demonstrations of assistance be used to improve the navigation performance of powered wheelchair users with hand-control disabilities? For this, we first contribute a custom teleoperation platform that enables demonstrations of assistance and uses haptic and virtual reality interfaces to facilitate the interpretation of raw sensor data. This platform is used to test claimed features of LAD, and we show that the technique can adapt to different hand-control impairments and also generates personalised assistive models. Furthermore, we explore the issue of model generalisation to physically different environments, which had not been investigated before. We show that this is a challenging problem and, to overcome it, propose solutions in terms of data collection and preprocessing, training and evaluation procedures, and learning algorithms. With these adaptations, we demonstrate that our model can provide useful assistance, even in previously unseen environments. To accelerate research in this field, we developed software that can simulate the full triadic interaction (assistant/robot/driver) concerning the application of LAD for robotic wheelchairs. This software permits the simulation of multiple disabilities and environments, and also allows people to take up the roles of the simulated driver and/or assistant. Finally, to keep our assumptions and developments in check, we conducted two experimental evaluations with humans, assessing how multimodal interfaces affect one's capability to infer the intention of a driver, and how different learning algorithms impact the generalisation and assistive performance of LAD.