PerMIS 2009 Plenary Addresses
Vice President for Research and Development
5D Robotics, Inc.
Title: Measuring the Benefits of Intelligent Behavior for Robotic Threat Detection
Abstract: For robotic applications in hazardous, critical environments, the intelligence needed to provide functional value (i.e. reduced time, increased probability of detection, increased hazard source localization accuracy) cannot be derived from a single behavior (such as obstacle avoidance, mapping, or mine detection). Rarely do we find an integrated suite of capabilities that is capable of accomplishing an end-to-end mission. Intelligence requires not simply behavior, but also the ability to use behaviors effectively towards a highly complex set of real-world, mission-level requirements. If the level of robot initiative and autonomy used in real-world missions is to increase, the underlying mechanisms for behavior composition and human interaction must also change.
Many approaches to creating behaviors as well as orchestrating them have been offered by the community including a variety of machine learning based techniques. These methods and algorithms are often highly elegant, formalized methods intended to streamline the development and testing methodologies. Unfortunately, these all too often fail to provide truly intelligent systems that provide value in the real world. Why is this?
One clue may be found if we consider biology. Is there anywhere in biology where we can find an elegant, formalized, understandable method for behavior composition? Functional intelligence may be, in part, derived from many interwoven heuristics for sequencing and interleaving behavior. In the brain these heuristics are learned over time through experience and perhaps not in an elegant fashion. Artificial Neural Networks (ANNs) are intended to model the behavior derivations we find in biology, but although ANNs allow us to effectively capture particular perceptual and action pairings, we are still left with the fundamental problem of how to sequence and compose behaviors to get a real job done. Without this behavior composition, we may have capability, but enjoy meager intelligence.
Although this talk will not submit a solution to this fundamental challenge, I would like to share a variety of experiments which, over the past few years, have allowed us to metric various components of intelligence for mobile robots used in a variety of real world missions. These missions include chemical plume localization, radiological characterization, urban search and rescue, mine detection and defeat of improvised explosive devices. To accomplish end-to-end missions in the hands of operators with no or little experience with robots requires a means to fuse components of robot intelligence while hiding the behavioral complexity from the user.
The Robot Intelligence Kernel (RIK) is being used to coalesce software components for perception, communication, behavior, world modeling, and human interaction into a single behavior architecture that can be easily transferred for use with a wide variety of robots and sensor-suited, low-level proprietary controls. This talk will discuss implementation strategies employed to integrate these components into a functional system that provides high-performance utility for various real-world tasks. Of particular interest is the cognitive glue, a fuzzy logic rule base, used to sequence and blend these behaviors into mission-level capabilities, such as minesweeping or radiological characterization. Lastly, the paper discusses agents within the interface that fuse various forms of robot and world representation. The interface agents also filter and interpret human input in order to incorporate it seamlessly into the behavioral intelligence of the robotic system. Our strategy is to hide sensor and behavior complexity while providing a means to integrate human intelligence at an appropriate level. In reviewing the benefits and limitations of the RIK approach, the talk will provide system-oriented results from recent hazard detection experiments. In particular, the talk will detail a number of measurements focused on the complete (i.e. human + robot + software + interface) system metrics as well as various component measurements.
Biography: Mr. David J. Bruemmer is Vice President for Research and Development at 5D Robotics, Inc where he is also a founder and board member. Prior to joining 5D Mr. Bruemmer was Technical Director for Unmanned Vehicles at the Idaho National Laboratory (INL.) For more than 14 years Mr. Bruemmer has enjoyed finding ways to fuse emerging science and engineering into innovative technologies that can change the way robots interact with humans and their environment. He has authored over 50 peer reviewed journal articles, book chapters and conference papers in the area of intelligent robotics. Mr. Bruemmer has been recognized by the President's Office of Science and Technology Policy for his work to forge effective interagency research collaborations across the Federal government (e.g. NASA, Dept. of Energy, Dept. of Defense, Dept. of Commerce, Dept. of Homeland Defense). He is a winner of the R & D 100 Award, the Stoel Reeves Idaho Innovation Award and the Federal Lab Consortium Award for Excellence in Technology Transfer.
The Robot Intelligence Kernel (RIK), developed by Mr. Bruemmer and his team, is being used as a framework for integrating robot software into a standardized, interoperable architecture. Mr. Bruemmer has developed robot behaviors used for a wide variety of robots for applications including remote characterization of high radiation environments, mine sweeping operations, military reconnaissance, IED defeat, chemical plume tracing and search and rescue operations. These efforts have yielded 11 Patents (Issued and Pending) and 10 copyrighted software inventions. His research in the area of countermine operations has demonstrated a four fold decrease in time necessary to find landmines and an improvement of over 20% in probability of detection when compared with the current military baseline. Before working at the INL, Mr. Bruemmer served as a consultant to the Defense Advanced Research Projects Agency, where he worked to coordinate development of autonomous robotics technologies across several offices and programs.
Prof. Paul Cohen
Professor and Head
Department of Computer Science
University of Arizona
Title: Against Sophistication: Why Worry About Performance Assessment
Abstract: The theme of the 2009 PerMIS is, "Does performance measurement accelerate the pace of advancement for intelligent systems?" Surely, performance measurement is necessary but not sufficient for the advancement of intelligent systems, and no measurement can compensate for badly designed performance tasks or for performance becoming an end in itself. AI is drunk on performing hard tasks at high levels. Given a choice between power and generality, most of us choose power. Our programs depend on designed exploits, or on designed search spaces in which programs can learn exploits. Divide-and-conquer, specific function, power over generality, and exploits are valuable engineering methods in many disciplines. They are apt to build machines that do one thing well. Human intelligence isn't that kind of machine.
Fixing the current situation will require a disciplined stand against sophistication. It will require investments in general, child-like intelligence, and the investors might not see a return—high performance from cognitive systems—for some time. I think this is a deal worth making, both because it is likely to succeed and because the pursuit of high performance returns low dividends.
Biography: Paul Cohen is Professor and Head of Computer Science at the University of Arizona. Before that he worked at UMass Amherst and the USC Information Sciences Institute. His research is on planning, learning, cognitive development and language. He wrote a textbook on empirical methods for computer science and has worked on the evaluations of several DARPA programs, most recently PAL, Coordinators and Machine Reading.
Prof. Raffaello D'Andrea
Measurement and Control Laboratory
Department of Mechanical and Process Engineering
Title: Towards a Ten Thousand Mobile Robot Warehouse
Abstract: Order fulfillment is a multi-billion dollar business. Existing solutions range from the highly automated, whose cost effectiveness is inversely related to their flexibility, to people pushing carts around in warehouses manually filling orders, which is very flexible but not very cost effective. In this talk I will describe a radical new approach to order fulfillment that is both flexible and cost effective. The key idea is to use hundreds of networked, autonomous mobile robots that carry inventory-storing pods to human operators. The result is a distribution facility that is dynamic, self-organizing, and adaptive.
Various challenges had to be overcome in order to make this an economically viable system, ranging from design of robust autonomous mobile robots, real-time wireless control of hundreds of moving agents, the coordination of these agents, and the design of various algorithms that allow the system to adapt and reconfigure itself based on the environment and operating conditions. I will discuss these challenges and how they scale to future warehouses with thousands—not just hundreds—of mobile robots.
Biography: Raffaello D'Andrea received the B.Sc. degree in Engineering Science from the University of Toronto in 1991, and the M.S. and Ph.D. degrees in Electrical Engineering from the California Institute of Technology in 1992 and 1997. He was an assistant, and then an associate, professor at Cornell University from 1997 to 2007. He is currently a full professor of automatic control at ETH Zurich. He is also a founder of, and chief scientific advisor for, Kiva Systems.
He is a co-recipient of the 2008 IEEE/IFR Invention and Entrepreneurship Award, a United States Presidential Early Career Award for Science and Engineering, and was the faculty advisor and system architect of the Cornell Robot Soccer Team, four-time world champions at the international RoboCup competition in Sweden, Australia, Italy, and Japan. He is a recipient of two best paper awards from the American Automatic Control Council and the IEEE, a National Science Foundation Career Award, and several teaching awards in the area of project-based learning. A creator of dynamic sculpture, his work has appeared at various international venues, including the National Gallery of Canada, the Venice Biennale, the Luminato Festival, Ars Electronica, and ideaCity.
Prof. Ben Kuipers
Computer Science and Engineering Division
University of Michigan, Ann Arbor
Title: Evaluating the Robot Cognitive Mapper
Abstract: A robot observes the space within range of its sensors. In this "small-scale" space, it detects hazards and makes local motion plans. As it explores its global environment, it knits local spatial models together to build a cognitive map --- a representation of the global structure of "large-scale" space that extends beyond the sensory horizon of the robot at any given time.
We have developed the Hybrid Spatial Semantic Hierarchy (HSSH), a model of the cognitive map that covers both large-scale and small-scale space, as experienced by the exploring robot. The key idea behind the HSSH is to combine the strengths of multiple different representations (ontologies) for space, each relatively simple: the Local Metrical, Local Topological, Global Topological, and Global Metrical maps.
This hierarchy of representations supports a relatively simple and robust way for the robot to construct a useful cognitive map from exploration experience. It also supports robust and efficient planning of routes from one place to another, as well as multiple ontologies for communication between a robot and a human directing it in how to reach a desired destination. The structure of the HSSH allows us to factor the evaluation task into simpler elements. Each level of the hierarchy can be evaluated according to its ability to meet the needs of the other levels, and the hierarchy as a whole is evaluated according to the different ways it can meet the needs of the robot agent, and how well each of those ways is accomplished. As a result of this factoring, each component is easier to evaluate, and has a lower bar for successful performance.
Biography: Benjamin Kuipers joined the University of Michigan in January 2009 as Professor of Computer Science and Engineering. Prior to that, he held an endowed Professorship in Computer Sciences at the University of Texas at Austin. He received his B.A. from Swarthmore College, and his Ph.D. from MIT. He investigates the representation of commonsense and expert knowledge, with particular emphasis on the effective use of incomplete knowledge. His research accomplishments include developing the TOUR model of spatial knowledge in the cognitive map, the QSIM algorithm for qualitative simulation, the Algernon system for knowledge representation, and the Spatial Semantic Hierarchy model of knowledge for robot exploration and mapping. He has served as Department Chair at UT Austin, and is a Fellow of AAAI and IEEE.
Prof. Tom Mitchell
E. Fredkin University
Professor and Department Head
Machine Learning Department
Carnegie Mellon University
Title: How does Brain Activity Represent Word Meanings?
Abstract: How does the human brain represent meanings of words and pictures in terms of the underlying neural activity? This talk will present our research using machine learning methods together with fMRI brain imaging to study this question. One line of our research has involved training classifiers that identify which word a person is thinking about, based on their neural activity observed using fMRI. A more recent line involves developing a computational model that predicts the neural activity associated with arbitrary English words, including words for which we do not yet have brain image data. Once trained, the model predicts fMRI activation for any other concrete noun appearing in the text corpus, with highly significant accuracies over the 100 nouns for which we currently have fMRI data. See a video from a recent CBS 60 Minutes story on Professor Mitchell's work "Reading your Mind."
Biography: Tom M. Mitchell is the E. Fredkin University Professor and head of the Machine Learning Department at Carnegie Mellon University. Mitchell is a past President of the American Association of Artificial Intelligence (AAAI), and a Fellow of the AAAS and of the AAAI. His general research interests lie in machine learning, artificial intelligence, and cognitive neuroscience. Mitchell's web home page is www.cs.cmu.edu/~tom.
Dr. Lora Weiss
Georgia Institute of Technology
Title: Assessing Autonomous Systems As They Evolve
Abstract: Today, unmanned systems are operating in-theater with untested collaborative capabilities. The vehicles are heterogeneous, in that they are developed by different contractors, they have different levels of autonomy, they have different sensors and capabilities, and they are physically disparate. Unmanned air vehicles built by one contractor have never autonomously collaborated with unmanned sea surface vehicles built by another contractor, and no one knows how they would perform if deployed together today. Their integrated use, however, is rapidly growing in the military. As improvements in autonomy, sensing, and reasoning advance, collaborating, multi-vendor unmanned systems will be increasingly employed to support challenging, tactical operations. The anticipated increase in sophistication drives the need for an ability to robustly test, measure, and evaluate heterogeneous unmanned vehicles for full spectrum dominance and joint operations. We need to consider assessment methods to evaluate force-on-force and mission level the effectiveness of disparate unmanned systems collaborating in theater-wide scenarios. A key requirement for assessing autonomous unmanned systems is the realization that unmanned vehicles pose new challenges that are distinct from traditional approaches to assessing systems. These challenges stem from the upcoming capabilities of unmanned systems being able to autonomously collect and process data, turn it into valued information and knowledge, and then intelligently act upon it with little to no operator involvement. Autonomy at the individual vehicle level involves transitioning cognition into decisions that drive actions. Based on the mission or operational environment, these unmanned systems may execute behaviors that cannot be precisely predicted. Assessments need to support evaluation of autonomous vehicle actions and judge whether the actions are reasonable and acceptable, without having precisely quantifiable metrics. Evaluating these systems will focus more on capabilities and missions rather than mechanics. New approaches to measuring their effectiveness will be adopted to support advances in autonomy and cognition, where the metrics and methods evolve and adapt, just as the systems do.
Biography: Dr. Lora G. Weiss is a lab Chief Scientist at the Georgia Tech Research Institute, where she conducts research on the design, development, and implementation of autonomy and control for manned and unmanned systems. She has supported intelligent autonomy for unmanned underwater vehicles, unmanned air vehicles, and unmanned ground vehicles, and is currently engaged in research in exploring all aspects of the behavior of these systems. Dr. Weiss has chaired sessions at IEEE conferences, ASA conferences, and Navy Symposiums and currently chairs the ASTM Standards Development Subcommittee F41.01, on Unmanned Maritime Vehicle Autonomy and Control. Dr. Weiss is on the Board of Directors for AUVSI, the world's largest non-profit unmanned systems organization. She has developed a video for IEEE Educational Services and has received several publication awards. Dr. Weiss has been Principal Investigator on numerous DoD programs sponsored by offices such as DARPA, the Office of Naval Research, and various Navy Program Executive Offices. She has provided over 150 technical briefs to high-ranking DoD officers and DoD technology offices.