Sensorimotor Cross-Behavior Knowledge Transfer for Grounded Category Recognition

Abstract

Humans use exploratory behaviors coupled with multi-modal perception to learn about the objects around them. Research in robotics has shown that robots too can use such behaviors (e.g., grasping, pushing, shaking) to infer object properties that cannot always be detected using visual input alone. However, such learned representations are specific to each individual robot and cannot be directly transferred to another robot with different actions, sensors, and morphology. To address this challenge, we propose a framework for knowledge transfer across different behaviors and modalities that enables a source robot to transfer knowledge about objects to a target robot that has never interacted with them. The intuition behind our approach is that if two robots interact with a shared set of objects, the produced sensory data can be used to learn a mapping between the two robots' feature spaces. We evaluate the framework on a category recognition task using a dataset containing 9 robot behaviors performed multiple times on a set of 100 objects. The results show that the proposed framework can enable a target robot to perform category recognition on a set of novel objects and categories without the need to physically interact with the objects to learn the categorization model.