by Paul Scharre
Kumar and Lee aren’t weapons designers, so it may not be at the forefront of their minds, but it’s worth pointing out that the technologies FLA is building aren’t even the critical ones for autonomous weapons. Certainly, fast-moving quadcopters could have a variety of applications. Putting a gun or bomb on an FLA-empowered quadcopter isn’t enough to make it an autonomous weapon, however. It would still need the ability to find targets on its own. Depending on the intended target, that may not be particularly complicated, but at any rate that’s a separate technology. All FLA is doing is making quadcopters maneuver faster indoors. Depending on one’s perspective, that could be cool or could be menacing, but either way FLA doesn’t have anything more to do with autonomous weapons than self-driving cars do.
DARPA’s description of FLA didn’t seem to stack up against Stuart Russell’s criticism. He has written that FLA and another DARPA program “foreshadow planned uses of [lethal autonomous weapon systems].” I first met Russell on the sidelines of a panel we both spoke on at the United Nations meetings on autonomous weapons in 2015. We’ve had many discussions on autonomous weapons since then and I’ve always found him to be thoughtful, unsurprising given his prominence in his field. So I reached out to Russell to better understand his concerns. He acknowledged that FLA wasn’t “cleanly directed only at autonomous weapon capability,” but he saw it as a stepping stone toward something truly terrifying.
FLA is different from projects like the X-47B, J-UCAS, or LRASM, which are designed to engage highly sophisticated adversaries. Russell has a very different kind of autonomous weapon in mind, a swarm of millions of small, fast-moving antipersonnel drones that could wipe out an entire urban population. Russell described these lethal drones used en masse as a kind of “weapon of mass destruction.” He explained, “You can make small, lethal quadcopters an inch in diameter and pack several million of them into a truck and launch them with relatively simple software and they don’t have to be particularly effective. If 25 percent of them reach a target, that’s plenty.” Used in this way, even small autonomous weapons could devastate a population.
There’s nothing to indicate that FLA is aimed at developing the kind of people-hunting weapon Russell describes, something he acknowledges. Nevertheless, he sees indoor navigation as laying the building blocks toward antipersonnel autonomous weapons. “It’s certainly one of the things you’d like to do if you were wanting to develop autonomous weapons,” he said.
It’s worth nothing that Russell isn’t opposed to the military as a whole or even military investments in AI or autonomy in general. He said that some of his own AI research is funded by the Department of Defense, but he only takes money for basic research, not weapons. Even a program like FLA that isn’t specifically aimed at weapons still gives Russell pause, however. As a researcher, he said, it’s something that he would “certainly think twice” about working on.
WEAPONS THAT HUNT IN PACKS: COLLABORATIVE OPERATIONS IN DENIED ENVIRONMENTS
Russell also raised concerns about another DARPA program: Collaborative Operations in Denied Environments (CODE). According to DARPA’s official description, CODE’s purpose is to develop “collaborative autonomy—the capability of groups of [unmanned aircraft systems] to work together under a single person’s supervisory control.” In a press release, CODE’s program manager, Jean-Charles Ledé, described the project more colorfully as enabling drones to work together “just as wolves hunt in coordinated packs with minimal communication.”
The image of drones hunting in packs like wolves might be a little unsettling to some. Ledé clarified that the drones would remain under the supervision of a human: “multiple CODE-enabled unmanned aircraft would collaborate to find, track, identify and engage targets, all under the command of a single human mission supervisor.” Graphics on DARPA’s website depicting how CODE might work show communications relay drones linking the drone pack back to a manned aircraft removed from the edge of the battlespace. So, in theory, a human would be in the loop.
CODE is designed for “contested electromagnetic environments,” however, where “bandwidth limitations and communications disruptions” are likely to occur. The means that the communications link to the human-inhabited aircraft might be limited or might not work at all. CODE aims to overcome these challenges by giving drones greater intelligence and autonomy so that they can operate with minimal supervision. Cooperative behavior is central to this concept. With cooperative behavior, one person can tell a group of drones to achieve a goal, and the drones can divvy up tasks on their own.
In CODE, the drone team finds and engages “mobile or rapidly relocatable targets,” that is, targets whose locations cannot be specified in advance by a human operator. If there is a communications link to a human, then the human could authorize targets for engagement once CODE air vehicles find them. Communications are challenging in contested electromagnetic environments, but not impossible. U.S. fifth-generation fighter aircraft use low probability of intercept / low probability of detection (LPI/LPD) methods of communicating stealthily inside enemy air space. While these communications links are limited in range and bandwidth, they do exist. According to CODE’s technical specifications, developers should count on no more than 50 kilobits per second of communications back to the human commander, essentially the same as a 56K dial-up modem circa 1997.
Keeping a human in the loop via a connection on par with a dial-up modem would be a significant change from today, where drones stream back high-definition full-motion video. How much bandwidth is required for a human to authorize targets? Not much, in fact. The human brain is extremely good at object recognition and can recognize objects even in relatively low resolution images. Snapshots of military objects and the surrounding area on the order of 10 to 20 kilobytes in size may be fuzzy to the human eye, but are still of sufficiently high resolution that an untrained person can discern trucks or military vehicles. A 50 kilobit per second connection could transmit one image of this size every two to three seconds (1 kilobyte = 8 kilobits). This would allow the CODE air vehicles to identify potential targets and send them back to a human supervisor who would approve (or disapprove) each specific target before attack.
But is this what CODE intends? CODE’s public description explains that the aircraft will operate “under a single person’s supervisory control,” but does not specify that the human would need to approve each target before engagement. As is the case with all of the systems encountered so far, from thermostats to next-generation weapons, the key is which tasks are being performed by the human and which by the machine. Publicly available information on CODE presents a mixed picture.
A May 2016 video released online of the human-machine interface for CODE shows a human authorizing each specific individual target. The human doesn’t directly control the air vehicles. The human operator commands four groups of air vehicles, labeled Aces, Badger, Cobra, and Disco groups. The groups, each composed of two to four air vehicles, are given high-level commands such as “orbit here” or “follow this route.” Then the vehicles coordinate among themselves to accomplish the task.
Disco Group is sent on a search and destroy mission: “Disco Group search and destroy all [anti-aircraft artillery] in this area.” The human operator sketches a box with his cursor and the vehicles in Disco Group move into the box. “Disco Group conducting search and destroy at Area One,” the computer confirms.
As the air vehicles in Disco Group find suspected enemy targets, they cue up their recommended classification to the human for confirmation. The human clicks “Confirm SCUD” and “Confirm AAA” [antiaircraft artillery] on the interface. But confirmation does not mean approval to fire. A few seconds later, a beeping tone indicates that Disco Group has drawn up a strike plan on a target and is seeking approval. Disco Group has 90 percent confidence it has found an SA-12 surface-to-air missile system and includes a photo for confirmation. The human clicks on the strike plan for more details. Beneath the picture of the SA-12 is a small d
iagram showing estimated collateral damage. A brown splotch surrounds the target, showing potential damage to anything in the vicinity. Just outside of the splotch is a hospital, but it is outside of the anticipated area of collateral damage. The human clicks “Yes” to approve the engagement. In this video, a human is clearly in the loop. Many tasks are automated, but a human approves each specific engagement.
In other public information, however, CODE seems to leave the door open to removing the human from the loop. A different video shows two teams of air vehicles, Team A and Team B, sent to engage a surface-to-air missile. As in the LRASM video, the specific target is identified by a human ahead of time, who then launches the missiles to take it out. Similar to LRASM, the air vehicles maneuver around pop-up threats, although this time the air vehicles work cooperatively, sharing navigation and sensor data while in flight. As they maneuver to their target, something unexpected happens: a “critical pop-up target” emerges. It isn’t their primary target, but destroying it is a high priority. Team A reprioritizes to engage the pop-up target while Team B continues to the primary target. The video makes clear this occurs under the supervision of the human commander. This implies a different type of human-machine relationship, though, than the earlier CODE video. In this one, instead of the human being in the loop, the human is on the loop, at least for pop-up threats. For their primary target, they operate in a semiautonomous fashion. The human chose the primary target. But when a pop-up threat emerges, the missiles have the authority to operate as supervised autonomous weapons. They don’t need to ask additional permission to take out the target. Like a quarterback calling an audible at the scrimmage line to adapt to the defense, they have the freedom to adapt to unexpected situations that arise. The human operator is like the coach standing on the sidelines—able to call a time-out to intervene, but otherwise merely supervising the action.
DARPA’s description of CODE online seems to show a similar flexibility for whether the human or air vehicles themselves approve targets. The CODE website says: “Using collaborative autonomy, CODE-enabled unmanned aircraft would find targets and engage them as appropriate under established rules of engagement . . . and adapt to dynamic situations such as . . . the emergence of unanticipated threats.” This appears to leave the door open to autonomous weapons that would find and engage targets on their own.
The detailed technical description issued to developers provides additional information, but little clarity. DARPA explains that developers should:
Provide a concise but comprehensive targeting chipset so the mission commander can exercise appropriate levels of human judgment over the use of force or evaluate other options.
The specific wording used, “appropriate levels of human judgment,” may sound vague and squishy, but it isn’t accidental. This guidance directly quotes the official DoD policy on autonomy in weapons, DoD Directive 3000.09, which states:
Autonomous and semi-autonomous weapon systems shall be designed to allow commanders and operators to exercise appropriate levels of human judgment over the use of force.
Notably, that policy does not prohibit autonomous weapons. “Appropriate levels of human judgment” could include autonomous weapons. In fact, the DoD policy includes a path through which developers could seek approval to build and deploy autonomous weapons, with appropriate safeguards and testing, should they be desired.
At a minimum, then, CODE would seem to allow for the possibility of autonomous weapons. The aim of the project is not to build autonomous weapons necessarily. The aim is to enable collaborative autonomy. But in a contested electromagnetic environment where communications links to the human supervisor might be jammed, the program appears to allow for the possibility that the drones could be delegated the authority to engage pop-up threats on their own.
In fact, CODE even hints at one way that collaborative autonomy might aid in target identification. Program documents list one of the advantages of collaboration as “providing multi-modal sensors and diverse observation angles to improve target identification.” Historically, automatic target recognition (ATR) algorithms have not been good enough to trust with autonomous engagements. This poor quality of ATR algorithms could be compensated for by bringing together multiple different sensors to improve the confidence in target identification or by viewing a target from multiple angles, building a more complete picture. One of the CODE videos actually shows this, with air vehicles viewing the target from multiple directions and sharing data. Whether target identification could be improved enough to allow for autonomous engagements is unclear, but if CODE is successful, DoD will have to confront the question of whether to authorize autonomous weapons.
THE DEPARTMENT OF MAD SCIENTISTS
At the heart of many of these projects is the Defense Advanced Research Projects Agency (DARPA), or what writer Michael Belfiore called “the Department of Mad Scientists.” DARPA, originally called ARPA, the Advanced Research Projects Agency, was founded in 1958 by President Eisenhower in response to Sputnik. DARPA’s mission is to prevent “strategic surprise.” The United States was surprised and shaken by the Soviet Union’s launch of Sputnik. The small metal ball hurdling through space overhead was a wake-up call to the reality that the Soviet Union could now launch intercontinental ballistic missiles that could hit anywhere in the United States. In response, President Eisenhower created two organizations to develop breakthrough technologies, the National Aeronautics and Space Administration (NASA) and ARPA. While NASA had the mission of winning the space race, ARPA had a more fundamental mission of investing in high-risk, high-reward technologies so the United States would never again be surprised by a competitor.
To achieve its mission, DARPA has a unique culture and organization distinct from the rest of the military-industrial complex. DARPA only invests in projects that are “DARPA hard,” challenging technology problems that others might deem impossible. Sometimes, these bets don’t pan out. DARPA has a mantra of “fail fast” so that if projects fail, they do so before investing massive resources. Sometimes, however, these investments in game-changing technologies pay huge dividends. Over the past five decades, DARPA has time and again laid the seeds for disruptive technologies that have given the United States decisive advantages. Out of ARPA came ARPANET, an early computer network that later developed into the internet. DARPA helped develop basic technologies that underpin the global positioning system (GPS). DARPA funded the first-ever stealth combat aircraft, HAVE Blue, which led to the F-117 stealth fighter. And DARPA has consistently advanced the horizons of artificial intelligence and robotics.
DARPA rarely builds completed weapon systems. Its projects are small, focused efforts to solve extremely hard problems, such as CODE’s efforts to get air vehicles to collaborate autonomously. Stuart Russell said that he found these projects concerning because, from his perspective, they seemed to indicate that the United States was expecting to be in a position to deploy autonomous weapons at a future date. Was that, in fact, their intention, or was that simply an inevitability of the technology? If projects like CODE were successful, did DARPA intend to turn the key to full auto or was the intention to always keep a human in the loop?
It was clear that if I was going to understand the future of autonomous weapons, I would need to talk to DARPA.
5
INSIDE THE PUZZLE PALACE
IS THE PENTAGON BUILDING AUTONOMOUS WEAPONS?
DARPA sits in a nondescript office building in Ballston, Virginia, just a few miles from the Pentagon. From the outside, it doesn’t look like a “Department of Mad Scientists.” It looks like just another glass office building, with no hint of the wild-eyed ideas bubbling inside.
Once you’re inside DARPA’s spacious lobby, the organization’s gravitas takes hold. Above the visitors’ desk on the marble wall, raised metal letters that are both simple and futuristic announce: DEFENSE ADVANCED RESEARCH PROJECTS AGENCY. Nothing else. No motto or logo or shield. The organization’s confidence is apparent. The words s
eem to say, “the future is being made here.”
As I wait in the lobby, I watch a wall of video monitors announce DARPA’s latest project to go public: the awkwardly named Anti-Submarine Warfare (ASW) Continuous Trail Unmanned Vessel (ACTUV). The ship’s christened name, Sea Hunter, is catchier. The project is classic DARPA—not only game-changing, but paradigm-bending: the Sea Hunter is an entirely unmanned ship. Sleek and angular, it looks like something time-warped in from the future. With a long, narrow hull and two outriggers, the Sea Hunter carves the oceans like a three-pointed dagger, tracking enemy submarines. At the ship’s christening, Deputy Secretary of Defense Bob Work compared it to a Klingon Bird of Prey from Star Trek.
There are no weapons on board the Sea Hunter, for now. There should be no mistake, however: the Sea Hunter is a warship. Work called it a “fighting ship,” part of the Navy’s future “human machine collaborative battle fleet.” At $2 million apiece, the Sea Hunter is a fraction of the cost of a new $1.6-billion Arleigh Burke destroyer. The low price allows the Navy to purchase scores of the sub-hunting ships on the cheap. Work laid out his vision for flotillas of Sea Hunters roaming the seas: