更聪明的无人驾驶车

　　American technologists hope to improve the U.S. military’s unmanned ground vehicles (UGVs) in the coming years by giving them autonomous navigation, something that would require engineers to marry a variety of terrain-sensing devices with software powerful enough to make quick decisions.

　　The thousands of military robots used so far in Afghanistan and Iraq have been mostly remote-controlled, with human operators dictating almost everything they do, the exceptions being robots that can right themselves automatically when flipped, for example, or automatically backtrack if they lose radio contact.

　　Having smart, self-directed robots working in the service of ground soldiers has been a Defense Department dream for years. Advocates of autonomously navigating robots are hopeful that a decision by the Pentagon’s Joint Requirements Oversight Council to sign off on a futuristic, 25-year “Unmanned Systems Roadmap” for ground, air and sea robots will shore up political support for today’s development efforts and bring fresh research dollars.

　　Among the vehicles that could be improved by new sensor technologies are the Army’s forthcoming Humvee-sized Multifunction Utility/Logistics and Equipment (MULE) UGVs. Lockheed Martin is developing and testing versions of these six-wheeled vehicles under the Army’s Future Combat Systems networking program. One version would carry reconnaissance and surveillance sensors to find and destroy targets for dismounted infantry. The Army also plans to deploy a version to transport supplies and one to dig up mines. The first of 1,500 MULEs are scheduled to be capable of running by remote control or autonomously beginning in 2013 or 2014, said Don Nimblett, business development manager for Lockheed Martin Missiles and Fire Control’s Unmanned Systems Division. Another six-wheeled Lockheed vehicle, the smaller Squad Mission Support System, would carry gear for a single dismounted squad. It could be in the field as soon as early 2010, Nimblett said, with the possibility of “some serious distribution of systems,” perhaps ultimately numbering 4,000, in 2011-12.

　　In autonomous applications, these UGVs and their envisioned successors would depend on sensors and software as completely as we depend on our five senses and brain. But it is the power of software that is arguably most important to the future envisioned by the U.S. military in its 25-year robotics road map. Humans “have sensors better than what most animals have, but most animals have better sense-and-avoid capabilities than machines do,” said Robert Mandelbaum, program manager of the Defense Advance Project Research Agency (DARPA) learning locomotion program.

　　On unmanned vehicles, Mandelbaum said, more powerful software would fuse information from four main types of sensors: electro-optical cameras, radar, ladar and acoustic.

　　Acoustic sensors, which bounce ultrasound waves off objects and time their return to calculate distance, are rarely used in ground systems, Mandelbaum said. Electro-optical video cameras include visible-spectrum and infrared varieties, which detect heat — handy for night vision as well as for spotting warm-blooded beings and motors. Radar uses electromagnetic waves the same way acoustic sensors use sound. Then there’s ladar, the newest of the bunch and the one that’s had the greatest impact on recent advances in the push toward autonomous ground robotics. Ladar, or laser detection and ranging (also called lidar, for light detection and ranging) works on the same principle as radar and acoustic sensors, by firing laser pulses and timing their returns.

　　UGVs such as the MULE must be equipped with multiple sensors because no sensor can do it all. Optical systems have excellent spatial resolution and, operating in stereo, can use triangulation to calculate distance. But they can’t see through dust any better than humans. Triangulation has its range limits, at most 100 times the distance separating the eyes, Mandelbaum said. This means an optical system capable of triangulating a mile ahead would have to be 175 yards wide. Radar’s long wavelengths can see through dust and fog, but they have coarse resolution unless one builds large transmitters, he said.

　　Ladar provides excellent distance information, but its short wavelengths are foiled by snowstorms, fog and dust. Because each pixel involves firing a laser and capturing its information, its resolution is coarser than that of optical systems. Scanning ladar systems such as those made by Velodyne in the U.S. and SICK AG in Germany, which fire and capture sequentially, are also mechanically complex. At least for now, “there really doesn’t exist a production-worthy ground robotic ladar” hardened enough for military field use, said Steve DiAntonio, business development director for Carnegie Mellon University’s National Robotics Engineering Center.

　　Flash ladar, a new type of a solid-state ladar, is being watched closely in the UGV world, DiAntonio said. Rather than the sequential pulses of scanning ladar, it sends out bursts of light and collects them all at once. Flash ladar has high resolution, provides instantaneous images, and can be made solid-state and compact. But most have very short range. Advanced Scientific Concepts Inc., which might be the only company doing long-range flash ladar, reports it has a system capable of measurement to roughly 1-inch resolution from a mile away. But Roger Stettner, the company’s president, said the system had yet to be integrated into any production robots, and that, “I think there are more commercial than military applications.”

　　That is not to say there might not be some military uses. In April 2008, Massachusetts-based iRobot, one of the Future Combat Systems contractors, licensed Advance Scientific Concepts’ ladar technology. “Flash ladar provides immediate high-frequency data for optical avoidance and ranging, so there’s very little back-end processing involved,” said Jim Rymarcsuk, vice president of business development for iRobot’s government and industrial business. Plus, he said, flash ladars have the potential to be tough, small and relatively cheap. He said he sees them being in the field in a year or two, most likely on smaller robots.

　　IRobot’s products won’t move as fast as MULEs, making the shorter-range flash ladar less of an issue. But the sensors can’t be too costly. “The cost point is much more critical for us. You don’t want a $100,000 sensor nobody wants to lose,” he said.

　　MULE and other pioneering large UGVs would, in addition to being able to navigate diverse terrain between programmed GPS waypoints, accept teleoperation and remote control. Engineers at the Army’s Tank-Automotive Research, Development and Engineering Center at the Detroit Arsenal in Michigan are testing soldier-controlled devices as well as autonomous performance in off-road terrain. The technology demonstrator is the 6.5-ton, six-wheeled “Crusher” chassis, one of two built for a DARPA program that ended in 2008.

　　Northrop Grumman Robotic Systems is working on the navigation system, which includes a sensor package with scanning ladar, optical systems and GPS, said Don Hall, the company’s lead engineer on the program. Ladar is the primary system for obstacle detection and avoidance, he said, with optical systems used mainly for providing feeds for operation by soldiers.

　　But the future, after years of DARPA-funded testing and contractor efforts, is likely to be in sensor fusion, in which intelligent software would combine simultaneous feeds from optical, ladar and other sensors to give UGVs a richer sense of awareness possible with just one type of sensor.

　　“The sensor fusion approach of taking stereo optical and ladar and applying learning technologies and other techniques has been the breakthrough in getting speeds up in off-road terrain,” said Jim Pippine, a System Planning Corp. manager who has worked on several DARPA UGV programs.

　　What ladar alone might perceive as 4-foot monoliths ahead, optics can clarify as being shrubs to be mowed over or boulders to be avoided, DiAntonio said. Building unmanned systems able to navigate natural terrain equally well has taken years of testing at Army bases in New York, Colorado and Texas, he said.

　　Such navigation takes a computer system with three key elements, DARPA’s Mandelbaum said. One is perception, which translates the raw signals coming in from sensors into usable information such as “ground vehicle 20 degrees off north approaching at 2.5 mph.” The second is internal representation, which uses sensor inputs to develop a 3-D map of the world around the UGV, and the third is a planning module for the UGV to decide where it should go next. It’s in such processing that the big challenges remain, Mandelbaum said.

　　Pippine said today’s technology is already relatively mature as far as navigating off-road and even on-road environments with other vehicles merging and following traffic laws, as was the case with the 2007 DARPA Urban Challenge. It’s unstructured environments that present problems, he said.

　　Take a Baghdad bazaar as an example, Mandelbaum said — chaotic, unstructured and dynamic, with carts, bicycles and people moving in every direction. “The algorithms were not designed to deal with that,” he said.

　　Navigating such environments without a human touch would require a leap in algorithmic representation of such nebulous things as human intent, quickly sizing up an environment for good guys, bad guys and “agnostics,” as Mandelbaum put it. Whether the limiting factors in such an effort would be in computer hardware or software, Mandelbaum can’t say.

　　“We don’t know how to do the processing, so we don’t know how much processor we would need,” he said.