To Ask or Not To Ask: Robot-assisted Bite Acquisition with Human-in-the-loop Contextual Bandits
Keywords: Contextual Bandits, Bite Acquisition, Human-in-the-Loop Autonomy
TL;DR: We propose a human-in-the-loop contextual bandit algorithm for robot-assisted bite acquisition that trades off improved task performance and cognitive workload.
Abstract: Robot-assisted bite acquisition involves picking up food items that vary considerably in their shape, size, texture, and compliance. An effective bite acquisition system should be able to generalize to out-of-distribution instances in a few-shot manner, but this is difficult for a fully autonomous strategy due to the large variety of food items that exist. In this work, we propose a contextual bandit algorithm that asks for human feedback to improve generalization to novel food items, while minimizing the cognitive workload that querying imposes on the human. We demonstrate through experimentation on a dataset of 16 food items that our algorithm improves the tradeoff between task performance and cognitive workload compared to two baselines: (1) a state-of-the-art fully autonomous baseline, and (2) a naive querying algorithm that does not incorporate cognitive workload.
Submission Number: 31
Loading