Multimodal embodied attribute learning by robots for object-centric action policies

Robots frequently need to perceive object attributes, such as red, heavy, and empty, using multimodal exploratory behaviors, such as look, lift, and shake. One possible way for robots to do so is to learn a classifier for each perceivable attribute given an exploratory behavior. Once the attribute classifiers are learned, they can be used by robots to select actions and identify attributes of new objects, answering questions, such as "Is this objectred andempty ?" In this article, we introduce a robot interactive perception problem, called Multimodal Embodied Attribute Learning (meal), and explore solutions to this new problem. Under different assumptions, there are two classes of meal problems. offline-meal problems are defined in this article as learning attribute classifiers from pre-collected data, and sequencing actions towards attribute identification under the challenging trade-off between information gains and exploration action costs. For this purpose, we introduce Mixed Observability Robot Control (morc), an algorithm for offline-meal problems, that dynamically constructs both fully and partially observable components of the state for multimodal attribute identification of objects. We further investigate a more challenging class of meal problems, called online-meal, where the robot assumes no pre-collected data, and works on both attribute classification and attribute identification at the same time. Based on morc, we develop an algorithm called Information-Theoretic Reward Shaping (morc-itrs) that actively addresses the trade-off between exploration and exploitation in online-meal problems. morc and morc-itrs are evaluated in comparison with competitive meal baselines, and results demonstrate the superiority of our methods in learning efficiency and identification accuracy.