Feeding is an activity of daily living (ADL) that millions of people need assistance with. Robots have the potential to help feed care recipients, potentially providing them with more independence while reducing the caregiver burden. For people with severe mobility limitations who experience little-to-no head movement, a robotic feeding system must be able to safely, comfortably, and effectively transfer food items directly into one's mouth. This thesis demonstrates a feeding system that can perform real-time head estimation using a novel mouth-perception pipeline invariant to utensil occlusion. Additionally, the system uses multi-modal contact classification to distinguish different types of contacts, including ones that occur due to secondary medical conditions often seen in people with severe mobility limitations. With this classification, the system can switch controllers to keep users safe while feeding them comfortably and effectively. A user study with 15 participants evaluates the real-time head perception of the system and shows that users preferred it to one without its capabilities. A second user study design evaluates the physical contact-aware controller switching employed by the feeding system, while a third user study design for people with mobility limitations explores the technology acceptance of our system.