ARS: What fired your interest in robotics?
Parker: I have always been fascinated by how easily we humans function in the world to perform "simple" daily tasks, like walking, talking, and reasoning. We receive a constant bombardment of sensory data and, with little to no effort, we make sense of the environment around us. We all cope remarkably well with uncertainty, change, and unexpected situations. We learn and adapt to the world around us with no need for an explicit instruction manual. We communicate, collaborate, and negotiate with others to operate as a community. We alter the world as we see fit to improve our living and working environments. Considering all that people - and even animals - can do, it is obvious that we all are incredibly complex, efficient creatures. My interest in robotics is a natural extension of this fascination with human capabilities. I believe one of the best ways to appreciate and understand the intricacies of how intelligent creatures function is to try to re-create this intelligence in artificial creatures. And, the engineering side of me also enjoys the satisfaction of seeing robots physically performing tasks that I programmed them to do, interacting with each other and the environment.
ARS: How do you define the word "robot"?
Parker: Intuitively, I think of a "robot" as a mechanical device having a physical body, with sensors able to detect characteristics of the environment, mobility to be able to move within the environment, and reasoning capabilities to intelligently connect sensing with action. Further, I think of an "intelligent" robot as one that can operate in a non-static environment and deal with uncertainties in sensors and effectors. Some intelligent robots can also improve their performance through learning and adaptation.
ARS: You are currently involved in DARPA project. Could you briefly introduce the project? What are the aims of your part of the project?
Parker: Our project for DARPA's Software for Distributed Robotics (SDR) program is aimed at enabling large-scale heterogeneous mobile robot teams to cooperate to achieve surveillance and reconnaissance in indoor environments. The overall project involves mapping an indoor environment, detecting pre-identified objects of interest, deploying a mobile acoustic sensor network, and detecting intruder targets that move in the environment. The project has involved up to 100 robots of 3 primary types - one type of robot performs the mapping task, one type of robot is a "leader"-type robot to assist in the sensor net deployment, and the third type consists of the mobile sensor net robots. Our part of the project has been to achieve mobile sensor net deployment using leader robots, and the distributed acoustic sensor network. A key research challenge in this work is achieving fault tolerant heterogeneous teaming, in which robots share sensory information to collaboratively accomplish their goals. We successfully demonstrated these capabilities in a series of rigorously-controlled experiments conducted by DARPA at Ft. AP Hill, Virginia, in January of this year.
ARS: What is so fascinating in exploring robot teams?
Parker: Lots of things! Most realistic complex systems - whether natural or human-made - do not operate alone. Instead, they must interact with other systems and the environment in dynamic ways. This is a type of coordination - each system must make consideration of the other systems in the environment to ensure that interference is kept to a minimum. More interestingly, the capabilities of a single robot can be extended through collaboration and cooperation with other robots. Enabling robots to easily and efficiently interact with each other is another one of those capabilities that is easy for humans, but much more difficult for robots. I find this a fascinating challenge for engineering functional distributed robot systems. Multi-robot teams research is very rich in topics to explore, such as distributed motion control, action selection, multi-robot learning and adaptation, cooperative object transport and manipulation, reconfigurable robots, inter-robot communication, multi-robot mapping and exploration, human-robot teaming, mobile sensor networks, and so forth. There are enough interesting research topics to keep me, and the rest of the distributed robotics research community, busy for many years to come.
ARS: How can we exploit the emergent phenomena and self-organisation?
Parker: Since I am particularly interested in the practical, engineering-side of distributed robotics, I believe that one of the key requirements for exploiting emergent phenomena and self-organization is the ability to predict the boundaries of the possible emergent behaviors. In real-world systems, we usually have a specific mission to accomplish, with clear constraints on the solution. In these systems, we usually cannot afford uncertainty in the types of behaviors the distributed robot team will exhibit. But while emergent systems by definition are not precisely predictable, we do need an understanding of the boundaries of the behavior characteristics expected from the emergent global phenomenon in order to make use of these systems in practical applications. Techniques that allow us to make predictions about the global behavior of the distributed system are very important for us to be able to exploit these capabilities.
ARS: Why is learning important?
Parker: Our world is not static and is not fully predictable. We do not have uniform laws, similar to the laws of physics, that encapsulate concepts such as common sense. We don't have precise models of the interaction dynamics of robots with their environment. We can't read the minds of humans that share the environment. We don't know the system control definition of external agents in the robot's world. We can't precisely predict all the hardware degradations that the robot might experience. Simply put, we can't know or predict everything that a robot might encounter in performing its task. As the common adage states, only change is certain. So, to deal with all of these issues, robots must be able to learn and adapt to changes in their environment, in the robot team member capabilities, and in their task. Robot solutions without learning may operate correctly for a while, but they will eventually reach a state change from which they cannot recover. Learning and adaptation provide the opportunity to generate more robust solutions for practical applications.
ARS: What are the differences and main problems of applying learning in inherently cooperative distributed systems?
Parker: Inherently cooperative tasks are those that cannot be decomposed into independent subtasks to be individually assigned to different team members. Instead, the success of the team throughout its execution is measured by the combined parallel actions of the robot team, rather than individual robot actions. The main challenge of learning in these contexts is the challenge of credit assignment. In other words, in these domains it is difficult for a robot to distinguish the effects of its own actions from the effects of another teammates' actions. Thus, it is not immediately apparent how a single robot can learn on its own; instead, the learning effort is a cooperative exercise. The nature of these problems also leads to very large search spaces.
ARS: What learning method you use and why?
Parker: I don't believe that there is just one learning technique that is "best" for multi-robot cooperation. My collaborators and I have investigated several types of learning in multi-robot systems for different types of applications. These vary from adapting pre-defined system parameters to learning new cooperative behaviors. Techniques that we have explored include function learning, reinforcement learning, and neural network learning. From a practical perspective, my recent interests in multi-robot learning are in enabling a team that begins with effective behavior control to learn to deal with environmental state changes, and for robots to share their learning experiences with their teammates. I believe that this is an important practical need in real applications of multi-robot teams.
ARS: What is the hardest problem to be solved?
Parker: The hardest learning problem I can imagine is duplicating the learning process of humans, from infants to wise old sages. The ability to go from helpless creatures that are totally dependent upon their caretakers to learning about gravity, theory of mind, interpersonal relationships, common sense, and countless other aspects of this world is amazing. How can we begin to duplicate the breadth and depth of these capabilities in a single robot or teams of robots? This is an incredible challenge.
ARS: Could such problems be easier solved with artificial evolution?
Parker: I believe difficult, encompassing learning problems such as the human learning problem must involve a wide range of learning techniques. For some problems, artificial evolution would seem to be a reasonable approach. But other problems likely require different learning approaches.
ARS: If I'm a student thinking about starting to do research in the field, what can I do now to prepare?
Parker: One of the most important aspects of meaningful research is asking the right questions. It is important to understand the key unanswered questions in the field. To do this, the researcher needs to be familiar with the current state of the art in the area of interest, and the best way to do this is to read as much of the current research in the field as possible. Look in particular at the current assumptions of the research, and the proposed directions of future research. These are helpful in identifying open issues in the field. But in addition to being aware of the open research questions of the field, I always advise students to pursue the ideas that excite them, about which they have a true curiosity. Research involves a lot of hard work, and you want to be doing this research in an area that you are truly fascinated with. This natural curiosity will be a huge benefit in keeping your motivation strong while you put in the long days, months, and years into making your research contributions.
ARS: In year 2000 you have received the U.S. Presidential Early Career Award. What was the affect of this award on your scientific career?
Parker: This award is a great honor, and I feel very privileged to have been given this recognition. Awards like this validate my research and help raise the visibility of my work in several arenas. The publicity of this award at the local level led to the opportunity to speak to a number of community organizations that had been previously unfamiliar with my research. At the national and international levels, this award helped lead to my selection for various research "think-tanks", advisory boards, and working groups to focus on strategic planning and analysis of the field of robotics. More broadly, I believe this award encouraged me to consider broad-impact issues in my scientific research that can have wide-ranging positive impact to the field.
ARS: What are the most promising research areas in the field?
Parker: I believe the most promising research areas in multi-robot systems are the topics of fault tolerance, heterogeneous teaming, large-scale (100+) distributed robot teams, distributed learning and adaptation, and enabling robot teams to easily work with humans as peers or controllers.
Associate-professor, Department of Computer Science, University of Tennessee, USA
Published in: Volume 2, Number 2, June 2004
For further informations in German please click here!