Tuesday, May 18, 2010

Six issues in military robotics

Flight Global reported on 1 April that the Italian Air Force's new drone demonstration team—the Squadrilia Tranquilla—would be flying at the Farnborough Air Show this summer. I must say that I was pretty geared up about that, at least until my searches of the rest of the Internet failed to surface any other information on the outfit. Then I noticed that the name in the article of the squadron commander—Major Antonio "Pesce" da Prile—slightly rearranged as pesce d'aprile—means April Fool. (I confess that I missed until later how the photo credit had gone to one Charlie Croker—the main character in The Italian Job.)

It was all a clever joke, except that it was almost too clever by half. Much of what is described in the article is already technically quite conceivable, even if the concept doesn’t quite meet the safety standards of airshows yet. Even the doctored image of a formation flight by Alenia Sky-Xs seems hardly out of the ordinary. The believability of the gag is rather a testament to how far unmanned aviation has come in the past twenty years.

Tight formation flight according to predetermined paths may require local augmentation of GPS down to centimetric accuracies, but that’s already available. It may not quite be up to airshow safety standards yet, but that may not be long in coming. As the editors admit in owning up to the joke, “maybe a UAV flying display team isn't too far away!”

And yet, military robotics is clearly in its infancy as an industry. That tight formation flying is pretty, but if it really is along predetermined paths, it’s not so militarily useful. When drones can swarm with but modest operator involvement, we may have something. Appearing as a single radar image, for example, could be very handy. (That’s probably the one tactical lesson we all took from Top Gun back in ’86). As much progress as has been made, we’re not all quite there yet, and that points to the obvious questions of what  needs to be done.

Navigation, before the 1990s, was a rather diverse field of technological effort and operational art. Sometime after the widespread adoption of GPS, satellite and inertial techniques came to dominate the approaches to almost every navigation and positioning problem. The trouble, of course, is that geopositioning alone is not enough for militarily effective operations in close proximity to which is moving away from GPS in the research. It's thus notable how DARPA's recent but already legendary Urban Challenge still largely featured GPS-driven methods. For whatever limitations we may see in such heavy reliance on this single solution, its ubiquity and cost-effectiveness have suppressed interest in almost any alternative.

Until now. It is true that the exquistite solutions of, say, nuclear-tipped Cold War weapons, such as inertials of ICBM quality, are naturally exquisitely expensive. But progress is beginning to be made in lowering sensor costs by using multiple, lower-cost, and lower-power sources. Academic research into these multi-agent systems has become a big deal, and practical work is close behind. While GPS will remain ubiquitious, de novo efforts are now move into the electromagnetic spectrum.

Progress underwater remains restrained by an understandable lack of progess is sonar, which still lacks the fidelity needed for really novel autonomous robotic applications. Unmanned surface vessels aren't having quite the same problems, though the crowded waters navigation problem is still a significant one. It's rather easy to remotely fly a Predator through the largely friendly airspace over Afghanistan than to dispatch a hands-off robotic patrol boat through the Straits of Hormuz. Machine vision thus seems a field of tremendous promise, but one that is so difficult to master that it holds few real experts yet (to meet a few of them, go watch a robot soccer match). These sorts of capabilities could prove enormously helpful to autonomous systems in environments where bandwidth is constricted or communications are just plain down--a fear that has been vexing NATO planners for decades, even if the results are yet to be found on a real battlefield.

All these needs point to the problems of real world computing, which means getting robotics algorithms to run in finite time. Reinforcement learning has thus become a hot topic in academic research today--if robots can learn, they may need to think less, in a sense. One barrier standing in the way of more widespread military application of this approach is that a relatively high trial rate is required for finding the optimal policy in each tactical situation. Most tactical problems, however, are what physicists would recognize as few body problems--those with too many examples to treat as deeply analyzed cases, but too few to treat statistically.

Even if the computers can be found to do what's needed, power management remains a huge challenge. This, in practical terms, is what happens after you send the little guy out beyond the extension cord, but before his batteries conk out. It's alos a challenge against which infamously little progress has been made for decades. Batteries, to the disillusionment of the hybrid car set as well, just aren't all that impressive today. It's no wonder that innovative robotics builders are experimenting with solutions like solar arrays and (for the really new-school-but-old-school) autonomous sailing craft.

These piece parts are vital, of course, but even so, they are far from the whole picture. Holistic design, which is rather a more artful than mere systems engineering, is what's really required for the revolutionary progress that many of use think is possible. Just thinking about that helps illuminate how companies in the business should think about strategy. Here’s how Robin Murphy, the Raytheon professor of computer science and engineering at Texas A&M, put it on her blog back in September:

According to a good friend, Bill Kearns, at the turn of the last century, there were over 200 car manufacturers in North America. (His family’s business was one of them.) Each manufacturer had something special, a starter motor, independent suspension, what have you. An amazing array of advances, some redundant, many brilliant. But the problem was, they weren’t on the same car. Who wanted the latest, greatest engine on a car that you had to use with a hand crank? 

Guys like William Durant and Henry Ford are widely credited with mastering manufacturing, but that wasn't their full contribution. As Murphy put it, they didn't just master mass production, they mastered mass production of the right thing, as they were among the first entrepreneurs to view cars as more than just the sum of their parts. The superior technology of a component was not the reason for existence but rather a marketable feature of a desirable whole.

This observation rather reminds me of Apple’s approach: marketing the machine and its software as a synthetic whole, and not just a collection of feeds and speeds. It’s no wonder that the US Army sent a delegation to Cupertino last month to ask how to better integrate iPhones and their mobile applications into military operations—a process that is obviously well underway anyway. (My personal favorite for killer app is the sniper’s ballistic calculator on the iPod touch.) Reference to Apple is important because, as DARPA assets in its META program, personal electronics is but one of several industries that deal with complexity rather better than does the military aviation, or almost any other military-industrial line of work. So far, customers’ demands for simplicity in the robotics business are driving a meaningful response in the market. As Mitch Rosenberg, VP of marketing at Kiva Systems, put it last year, “robotics software is getting more complex, but succesful companies are shielding users from having to deal with that complexity.” It just doesn't do to make a plane harder to fly from Nevada than it is from inside a cockpit, and the industry actually figured that out well before its customers did.

Hence, some of the beauty of developing more advanced military robotics lies in the potential for removing complexity from some of those more exquisite military systems. Combat aircraft will be more useful in battle if they can be developed in months or years, and not decades. Removing the pilot from the cockpit, after all, does a lot to reduce the complexity of the airplane, while simultaneously improving some of  its performance parameters, such as range and endurance. Military customers are showing more interest in those potentially shorter product lifecycles (see the point above about DARPA), but they're absolutely seeking multi-mission flexibility at the same time. Consider how the US Army is openly thinking that its eventually replacements for the latest models of the Apache gunship and the Chinook heavy lift helicopters, probably debuting sometime after 2020, ought to be optionally-manned.

After all, as General David Deptula of the USAF reminded us on Vago Muradian's television whow back in December, the future looks quite unmanned in many areas, but not wholly so in any of them. Human-robot interaction for teaming and collaboration is thus another area ripe for advancement. A good R2 unit (try to keep a straight face) would have been valuable for NASA's dreams of missions to Mars, but more practically and immediately, could also make a great backseater in a fighter jet. Quite seriously, if aircraft can fly with pilots on the ground, and with multimember crews, there may be great potential in automatic more of the functions on single-seat aircraft, just to maximize the value of those eyes on the tactical situation.

Clearly this is great stuff, just as the immense progress of the past two decades has produced as well. And yet, that points to a sixth thrust for industry to consider, and one without engineering drawings or technical specifications. It's a matter of maintaining momentum. As Neal Blue of General Atomics commented a few weeks ago in his interview with Aviation Week magazine, American military acquisition always seems to eventually form committees to appoint other committees to oversee processes, and once that happens, the whole effort gets gummed up. The drive to reduce the natural failures of technical progress--widely mistaken for waste--leads to greater if longer-term waste, and with no benefit to the troops.

Bureaucratization is unlikely to improve performance in dynamic fields of endeavor. Plain and simple, this calls for old-fashioned lobbying, and even (gasp) the earmarking that funded such beloved advances as the Predator. Working around customers who don't know what they need or even want can be tricky and even risky, but as companies like GA have demonstrated, there can be real first-mover advantage in doing so. The outcomes will mean a lot to the troops, and the remuneration is richly deserved return on investment.

This piece can be found in its original entirety here.

No comments:

Post a Comment