Robot learning through observation via coarse-to-fine grained video summarization

Abstract

Learning human daily behavior is important for enabling robots to perform tasks and assist people. However, most prior work either requires specific sensors for capturing data or heavily relies on prior knowledge of human motion, which can be difficult to obtain. To alleviate the above problems, we propose a novel pipeline for robots to learn human behavior based on coarse-to-fine video summarization using a single Kinect camera. Specifically, the robot first retrieves information of general interest followed by a task-specific content retrieval, then focuses on fine-grained motion clips of human behavior, and guides itself by using an object-centric learning method to complete the desired task. Our work has three unique advantages: (1) it enables the robot to effectively capture granularity hierarchies of human behavior which efficiently exploits multi-stage information while alleviating disturbances and redundancies in visual data; (2) it obtains knowledge by focusing on object movements in summarized motion clips which does not require any prior knowledge of human motion; (3) it only requires a single Kinect sensor for the robot to learn human behavior which is fully accessible and easy to equip. Experiments in an office environment were performed to validate the efficiency and effectiveness of the proposed framework, and the results indicate that this approach exhibits good learning efficacy for the robot to understand human behavior and learn to perform tasks.