Quantifying Coherence when Learning Behaviors via Teleoperation Sekou Remy and Ayanna M. Howard Abstract— Applications of robotics are quickly changing. Just as computer use evolved from research purposes to everyday functions, applications of robotics are making a transition to mainstream usage. With this change in applications comes a change in the user base of robotics, and there is a pronounced move to reduce the complexity of robotic control. The move to reduce complexity is linked to the separation of the role of robot designer and robot operator. For many target applications, the operator of the robot needs to be able to correct and augment its capabilities. One method to enable this is learning from human data, which has already been successfully applied to robotics. We assert that this learning process is only viable when the demonstrated human behavior is coherent. In this work we test the hypothesis that quantifying the coherence in the provided instruction can provide useful information about the progress of the learning process. We discuss results from the application of this method to reactive behaviors. Such behaviors permit the learning process to be computationally tractable in real-time. These results support the hypothesis that coherence is important for this type of learning and also show that this property can be used to provide an avenue for self regulation of the learning process. I. I NTRODUCTION The description of the average robot user is changing. Just as computers have evolved from the realm of research and extreme novelty applications to becoming commonly found in multiple places in a typical modern home, as well as in the modern workplace - robotics is also making a similar transition. Changing applications results in a change in the typical user and whether an amputee, an arthritic grandmother, or a health care worker providing in-home assistance, a general theme is that people seek to have less computational effort expended to control the devices they use. Also, as the user base becomes more non-technical, there is an increased push to reduce the need to be technically involved in robotic instruction. With the change in demographics of the average robot user, there is a need to separate the role of robot designer and robot user. Such a separation would require that the user be able to provide new or fine tune existing robotic capabilities, and learning by teleoperation is one such method of accomplishing this task. For the changing user to utilize an in-home robot, we believe that this will allow the typical home user the ability to train the robot without placing heavy requirements on expertise and prior training. Learning from teleoperation can also be classified with similar approaches such as learning by demonstration [1], [2], and learning by Sekou Remy and Ayanna M. Howard are with the Human-Automation Systems (HumAnS) Lab, School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta GA 30332, USA {sekou, ayanna}@ece.gatech.edu observation [3], [4]. These methods have been applied to robotics and serve as a valid method of enabling a robot to perform new tasks. One challenge created by using such an approach is that the robot will need to be endowed with the ability to regulate when it learns in concert with its user/teacher/operator. We propose that this self regulation, a critical component of autonomous robotic learning from human supervision is hinged on some base properties that must be properly explored. One such property, coherence, has not been fully treated in robotics and in this work we seek to investigate coherence and its relationship to learning from human data. II. BACKGROUND There are two challenges with incorporating home users into the robotic learning process. First, the user may not be able to identify when the limits of the robot’s capabilities are reached and second, the user may not initially be capable of providing competent training to the robot. We believe that coherence is a property that can be used to address these two challenges. Coherence is a term that has been mentioned in several engineering and science domains. Merriam-Webster defines coherence as a “systematic or logical connection or consis- tency”, but in each domain the definition varies to some degree. In robotics the term has been used in [5], [6], [7] but even in these usages, a process to quantify coherence is not evident. In this work we use the definition that coherence in teleoperation data is the property that forces the action state to be linked to a specific sensory state for each behavior. Coherence rises from a causal relationship between the actions executed and the sensory evidence provided. Because of this relationship the actions executed should be logically consistent with the evidence provided. In this work, we present an approach which enables a robot to quantify coherence in the instruction which it is provided. For humans, it is often times obvious that if data are not coherent this will pose challenges for learning. We believe that if robots are so equipped to also identify this property, then it can have positive impact, especially for applications in which autonomous robot learning is useful. The approach utilizes a property known as the mean quantization error which will be defined more fully in section III. By isolating characteristics of this property we show how it can be used to augment the process of learning from teleoperation when possible and also to identify when it is not possible for the robot to learn in that manner. Proceedings of the 17th IEEE International Symposium on Robot and Human Interactive Communication, Technische Universität München, Munich, Germany, August 1-3, 2008 978-1-4244-2213-5/08/$25.00 ©2008 IEEE 471