Tractably Adaptable Food Manipulation for Robot-Assisted Feeding

Abstract

Assistive robots can empower those with mobility impairments, engendering feelings of independence. However, to reach the point of actual in-home use, they must manage the trade-off between adaptability, tractability, and comfort. On one hand is the non-stationary distribution of the environment and user preferences; a robot may need to explore to find the best action to take. On the other hand, excessive exploration can make for an uncomfortable experience of failures and unpredictable motion. This is particularly salient for intimate tasks like feeding. Here we focus on the particular problem of food acquisition, with metrics, system design, and assumptions informed by studies with people with upper spinal cord injuries. The problem can be mapped into the well-studied contextual bandit framework to enable online adaptation with theoretical guarantees on performance, though these scale with the size of both the context space and the action space. Both can be large: set by the variety of food users want. We show how we can leverage haptic information and human expertise to shrink both of these spaces, making this online adaptation tractable. Finally, we describe a complete, portable system that can be used for an extended in-home deployment.