Software for Autonomous Systems 1. Good Morning. It’s truly a pleasure to have an opportunity to speak to you today about Software for Autonomous Systems. For most of my career I have worked in the area of robotics and unmanned systems (including both unmanned ground vehicles and unmanned aerial vehicles), and am currently the Army’s senior, uniformed roboticist. Someone once asked me what it meant to be the Army’s senior roboticist? My response was that it was probably a bit like being an Admiral in the Swiss Navy. The title sounds impressive, but if you’re really going to get anywhere, you’re going to need the help of lots of other people…the kind of help that, hopefully, will come from some of you here, today. 2. Like most robotics people, my background and education are primarily grounded in mechatronics. However, it was this very fact, namely the experiences of actually building robotic systems which inculcated in me a profound appreciation for the pivotal role played by the real-time embedded software, especially for applications seeking to demonstrate a meaningful degree of real world autonomy. 3. The software for autonomous systems effort in ITO currently comprises two related, but also significantly distinct areas. These include the software-only, autonomous systems (Knowbots), and the software for physically embodied, autonomous systems (Robots). The term “Knowbots” refers to a class of software agents that perform various information services, especially those that are targeted on negotiation for resource allocation and customization. Knowbots range over cyberspace. The term Robots, on the other hand, refers to the full range of physically embodied, autonomous systems which, though physically instantiated, also require software-enabled functionality. Robots range over the physical world. 4. Knowbots address the fundamental challenges of making our cyber systems more capable, more useable, and more intuitive; first by making the software for these systems more responsive to user requirements, and second by relegating some of the boring and tedious tasks to modestly intelligent, software agents. Knowbots may also help extend our information infrastructure to one that reaches and empowers ordinary citizens, and thereby enable a new class of information appliances that may one day become part of everyday life. 5. Robots offer the potential to undertake missions that might otherwise be impossible, thus creating new operational capabilities. The Mars rover can collect samples of the red planet’s soil, while unmanned submersibles can perform extended, highly dangerous undersea missions at staggering depths. Robotic Ground Vehicles have entered the Chernobyl reactor facility, performing assessment and remediation tasks where no human can venture and survive. Robots can also enhance or extend existing capabilities. In the Balkans, Predator and Hunter UAVs have provided critical aerial reconnaissance and surveillance, while Panther, unmanned ground vehicles have helped protect soldiers from landmines in Bosnia. Lastly, there are opportunities to enhance military capabilities while simultaneously reducing costs. Examples range from Unmanned Combat Air Vehicles (UCAVs) at the high end, which can forego pilot-driven design costs, to tiny micro-robots at the low end, that might be cheaply produced and employed by the tens of thousands. 6. ITO currently has four programs in this area of Software for Autonomous Systems. These include one knowbot program, called Autonomous Negotiating Targets (ANTS), and two robot programs, called Mobile Autonomous Robot Software (MARS) and Software for Distributed Robotics (SDR), respectively. A fourth, and closely related program, Software Enabled Control (SEC), will be described shortly by my colleague, Dr. Helen Gill. 7. The vision of the ANTS program is to autonomously negotiate the assignment and customization of resources (such as weapons) to tasks (such as targets), and hence the name. Such systems have components that must communicate with peers, as well as with information and command concentrators at higher levels of situation or mission abstraction. These agents explicitly represent goals, values, and assessments of likelihood and assurance, and reason about those quantities and qualities in order to accomplish their mission, often within demanding, time-bounded constraints. Ultimately, the act of negotiating serves to explicitly generate actionable information that would be difficult (if not impossible) to otherwise obtain in a sufficiently, timely fashion. 8. PAUSE… Actually, despite the reference to negotiation, this statement really has nothing to do with ANTs. However, I now have a pretty good feel for how much time you all have spent reading airline magazines. 9. But seriously, here we’ve depicted one potential application of negotiation technology, which is central to the ANTs program. Special operations missions often require teams of manned rotocraft to fly extremely low, at high speeds, and in close formation. During these dangerous maneuvers, there may not always be adequate time for warning systems to alert the crew so as to avoid collisions, such as the Australian Army’s fatal collision of two Blackhawk helicopters in 1996. That mishap is depicted here. The Boeing Company is investigating whether knowbots could address this problem by continuously negotiating mutable protection zones, intervene when pilot error appears imminent, and provide crew members with timely information regarding both the loss of, and return of pilot control. 10. Virtually all free ranging mobile robots (that is, those that can perform complex tasks outside of structured environments) must rely upon a human operator to generate the synchronous control inputs needed to appropriately interact with the environment. Essentially, the human acts as the “remote pilot or driver”. While a few systems can demonstrate a measure of limited autonomy, such as waypoint navigation, even these systems must revert to remote control when confronting circumstances or conditions which were not adequately accounted for in advance. While current systems have given us a glimpse of the promise that robots hold, they suffer from several significant limitations: 1) Vulnerability to the vagaries of wireless communications, both from propagation anomalies as well as signal strength attenuation. 2) Operator performance degradation due to limited sensory feedback and communication latencies. 3) And finally, a general inability to scale to large formations of robots, due primarily to the lack of suitable spectrum bandwidth. 11. Consequently, the MARS program is pursuing software technologies to enable real world autonomy. To do this, the program has two primary goals: First, to develop software solutions that synthesize the desirable features of both deliberative (symbol-mediated) and reactive (sensor-mediated) control, as well as incorporate software-enabled adaptability (including the facility to learn). This means that the robot could not only respond appropriately (within its limitations) to unforeseen circumstances, but also improve its ability to do so as it gains experience. The second goal addresses the software creation challenge by developing those technologies that facilitate the incorporation of both explicit programming and learning- derived mechanisms in order to generate reusable, real-time, embedded software for autonomous robot applications. Note that if we intend for these robots to be able to free range over the real world, they must have a significant capacity to function in unpredictable, dynamic, and unstructured environments. 12. Predictability is important for all mission critical applications. In military robotics, this has typically motivated a deliberative, symbol-mediated control system. Usefulness is generally proportional to the degree to which both the task and the world can be accurately modeled, in advance. Robustness issues, conversely, motivate control systems that rely primarily on the robot’s sensors (also called reactive or sensor-mediated control). Such control strategies perform well for simple tasks, but generally do not scale well to complex tasks, nor are such systems readily re-programmable. Fully compatible data structures that simultaneously support these two features (predictability and robustness) have thus far proven elusive. Learning and adaptability further complicate matters by raising the somewhat thorny issue of unanticipated emergent behaviors. Those of you with children know EXACTLY what I mean. Also, the ability to explicitly handcraft all of the required software for robust autonomy has proven intractable, despite several heroic (and very costly) attempts to do just that. Conversely, at least in theory, all the software should be able to be generated, implicitly through some form of machine learning, but actual results have been disappointing. Our strategy here is to pursue several hybrid approaches, in an effort to realize the benefits of both approaches while mitigating the limitations of each. 13. More specifically, the MARS program is pursuing three alternative approaches, which are primarily distinguished by the nature and degree to which they employ a learning-based software solution. 1) Soft computing is a collection of software technologies that are inherently tolerant of imprecision. It is generally considered to be the confluence of several complementary methodologies including fuzzy logic, artificial neural networks, and probabilistic reasoning (including evolutionary (or genetic) algorithms, chaos theory and belief networks.) Here learning is largely parametric. 2) Robot shaping is a term derived from experimental psychology that deals with the so- called “shaping” of animal behavior via training. A human trainer provides domain and task knowledge, while the robot automatically compiles the taskings into procedural instructions. This approach relieves the human of the need to manage most low-level implementation details. 3) Imitative learning is an idea that was spawned by the Japanese progress in humanoid robots, and emphasizes the significance of the robot’s physical embodiment in determining the richness (and hence the potential utility) of the machine learning experience. It is the most learning intensive of the three approaches. 14. Consider the full scale, singly autonomous robot of the future. It may take many different forms, from a conventional (ground, air, or maritime) vehicle to the humanoid robot who services it. Such systems could operate in the most contaminated or otherwise dangerous environments, without the need for a remote human operator to monitor and control every movement. Because of this, many robots could work together in close proximity, without interfering with one other. Indeed, overall force size could be readily tailored (scaling up or down as needed), independent of most human-driven considerations. What’s more, many of the technological components to realize these systems are available now, or are at least in development. 15. For example, the Defense Department’s Joint Robotics Program has explored a number of conventional and special purpose ground robotics platforms that have demonstrated significant utility across a wide range of military applications. This platform, built by Robotics Systems Technology of Maryland, for example, is being evaluated for both tactical and physical security applications. 16. In recent years the Honda Motor Company stunned the world with the walking and stair-climbing abilities of its humanoid robot, the P2, and later by the more refined version, the P3, which is shown here. These state-of-the-art, anthropomorphic, mechatronic devices are expected to augment human workers for selected applications in Japanese factories within ten years. One of the significant benefits of this technology is that work spaces, tools, and other infrastructure need not be modified to accommodate a unique robotic form factor. Since these robots are human-like in form, they can be integrated into existing work areas in a measured, relatively transparent fashion, and with minimal disruption, as both the robots’ capabilities and the circumstances warrant. 17. Besides the very capable, singly autonomous robots just described, there exist other autonomous robot instantiations whereby cooperation is not just desirable, but rather is essential to successful mission accomplishment. As depicted here, large numbers of small robots would work together in order to achieve an overall mission goal, such as minefield remediation. 18. Significant mechatronic research progress had been made here as well. The small hexapod robot, Aerial, built by IS Robotics of Massachusetts, can operate in the turbulence of the surf zone, a challenging environment indeed. Particularly interesting is the robot’s ability to operate even when flipped completely over, and hence Aerial has proven itself to be an excellent research tool for adaptive control software. As depicted here, Aerial robots might be used to seek out mines in order to clear a beach landing site. These examples should not be taken to suggest that there is no requirement for additional hardware research. That’s hardly the case. Rather, the point is simply that we have made demonstrable progress in developing some of the essential, mechatronic pieces to the puzzle. However, just as with our most sophisticated manned platforms, such as tactical aircraft, the software technologies for composing suitable, real-time embedded software have not kept up, and thus this class of software remains a particularly vexing challenge. 19. The Software for Distributed Robots (SDR) program seeks to address this software challenge in the very small robot regime. One reason for our investment here is that extremely small, micro-miniature robots offer a unique opportunity to achieve dramatic economies of scale, if only one can get them to effectively work together, in a collective fashion. Nature offers us many such examples, from honey bees to soldier ants. In each case, the effectiveness of the whole exceeds the simple aggregation of the effectiveness of the individual members. Dick Urban’s Distributed Robotics program in MTO is addressing a number of the key enabling technologies, from unique forms of locomotion to power systems. That work is discussed elsewhere. To compliment those efforts, the SDR program in ITO is focussed on developing enabling, embedded software technologies for these tiny platforms. 20. Because of the extremely small form factor that characterizes these devices, resources (including inter-robot communications, on-board processing capacity, and electrical power) are extremely limited. Consequently, this program is focussed on three research issues: 1) control strategies, specifically those for the highly coordinated control of large numbers of extremely small robots 2) very light weight, power efficient, networking protocols 3) and various data processing strategies that can satisfactorily perform in the face of these severely restrictive, on-board computing constraints. 21. As I indicated earlier, we continue to make progress, and real robotic systems are performing useful missions today. For example, consider the domain of ground based systems (and I include small rotocraft in that class), examples of demonstrated robot mission capability include countermine operations, explosive device handling, reconnaissance and surveillance, as well as physical security and force protection operations. Enabled by significant progress in our ability to create the missing software, future systems will dramatically expand this capability envelope. 22. The ultimate goal of these efforts is to enable the capability to pervasively employ autonomous robots at will, across a wide range of applications, and across the entire spectrum of conflict. Obviously, this vision will require a great deal of work, and we will need your help and support. But I believe it’s worth the effort. Such a capability, in my view, represents a profound, asymmetric advantage to U.S. Forces. I’m convinced that the Navy’s Tomahawk Cruise Missile, the Air Force Predator UAV, and the Army’s Panther UGV collectively give us a concrete sense of the revolutionary potential of military robotic systems. Besides the software research work which I spoke of here, Dr. Gill’s program in Software Enabled Control is particularly relevant to systems requiring very high-speed, controller performance, both manned and unmanned. She’ll speak more on that shortly. Lastly, lest you misunderstand my passion for this research, I want to close with just a few personal thoughts. 23. Especially in the latter half of this century, America’s wars have often been very complex things, and they don’t seem to be getting any simpler. Academics will no doubt continue to chronicle their histories, while retired politicians (and even a few retired generals for that matter) will pen their memoirs. Students will endlessly study and ponder the tactics and strategies employed, long after the din of battle has faded. But with all this hubbub attending to our nation’s wars, the key figure is often forgotten; that one indispensable player on the stage of our nation’s history, the American soldier. Thank you.