Business & GA, Commercial

UAVs That Swarm

By Charlotte Adams | October 1, 2003
Send Feedback

If the command and control (C2) of today’s mammoth unmanned air vehicles (UAVs) seems like a tough problem, consider futuristic, bio-inspired concepts for autonomous UAV systems. Encouraged by studies of insect colony dynamics, researchers have churned out numerous papers and simulations to investigate how insect "swarming" behavior can be applied to the control of UAVs. The use of algorithms derived from insect behavior–and even the behavior of subatomic particles–promises the ability to control UAVs implicitly, from within the group, rather than explicitly from outside the group, to reduce bandwidth consumption, lower software development cost, and improve survivability.

Why is swarming technology important? Several reasons. Because swarms would be more self-sufficient, one human could look after more UAVs. The random motions of swarm members could render the UAVs more resistant to individual losses, by making them more difficult to target. Area searches, where there is no a priori information, and tasks such as surveillance and the suppression of enemy air defense also could be appropriate. One can imagine the possibilities for diversions and psychological warfare. Such possibilities are mentioned in a recent paper by Bruce Clough, technical area leader-control automation for the Air Vehicles Directorate of the Air Force Research Laboratory (AFRL). Clough’s own work, however, as described below, differs from "pure" swarming research.

Hitting the Mainstream

Swarming algorithms also are beginning to surface in the commercial world, in optimization tools for analyses of the stock market and futures markets, which can exhibit swarm-like behavior.

Swarming concepts involve the imitation of insect-like "emergent" behavior–group behavior that emerges from individuals’ interaction with their environment and each other. Because swarm behavior emerges from local interactions and uses local, ad hoc communications, less bandwidth would be required and control software code could be smaller and less expensive to develop.

Swarming research ranges from "pure" swarming, described above, to embedding emergent behavior algorithms within a more predictable control system. The use of more autonomous UAV swarms lies in the future, perhaps only when "the Nintendo generation gets stars on their shoulders," predicts David Scheidt, group scientist for the Information Technology Group of the Applied Physics Laboratory (APL). Scheidt’s group at APL is preparing to conduct two hardware-in-the-loop demonstrations of UAV and UAV/ground robot swarms.

Leaving an Ant Trail

The APL demos will apply a combination of "pheromonal" and "artificial physics" approaches to UAV control. The pheromone approach uses software processes which imitate the chemical messages that insects such as ants secrete. Ant pheromones mark trails from the nest to a food source–which they have found by purely random searches–and from the food source back to the nest. As more ants discover the trail and add their pheromones to it, it becomes like a superhighway while trails leading nowhere fade away.

In the world of UAVs, a vehicle could transmit a "digital pheromone," or message, to a neighboring craft in order to indicate what it has seen. The pheromones could be processed by the vehicles themselves or perhaps by hardware elements distributed across the battlefield. Internal processing becomes more difficult as the number of messages multiplies. But APL has found a more efficient way to process the information, Scheidt says.

APL also uses algorithms derived from mathematical descriptions of the behavior of subatomic particles. APL’s hybrid, artificial physics/pheromonal technique, arcane though it may sound, has yielded promising results. The lab has simulated groups of up to 200 air vehicles performing tracking and trailing, survey and mapping, formation flying, establishing a communications link between two entities, and calling in another vehicle for data fusion and target classification. Problems that crop up with artificial physics include the tendency of the simulated vehicles to jitter, or move–just as subatomic particles do–and to get stuck in unproductive "cul de sacs." APL’s hybrid approach, however, attempts to remedy the defects of either method alone.

Unlike most researchers in this field, APL will be able to take its work a step further, with actual hardware-in-the-loop demonstrations. The first, sponsored by the U.S. Joint Forces Command and expected next summer, involves applications like patrolling, detection, calling in other vehicles for object classification, tracking and formation flight, using approximately four UAVs. The second, internally funded demo probably will involve four unmanned air vehicles and 16 ground vehicles.

Applying ‘Microbehaviors’

Researchers at AFRL’s Air Vehicles Directorate, meanwhile, are applying emergent "microbehaviors" within the bounds of a more deliberate and predictable control system. The main focus is to use emergent behaviors to reduce the size and cost of control software for future large systems, says Clough. But these algorithms will be based on the emergent behavior of highly trained pilots operating in a manned strike package.

As Clough stated in a recent essay, the costs per line of flight-critical, higher-order code for vehicle management systems (VMS) in 2002 was "pushing $600" (see chart above). According to rough estimates, VMS code developed to support truly autonomous UAVs–using current software and hardware architectures–could balloon to between 1.8 and 4.5 million lines, which would add up to at least $900 million. "We would find ways of getting the cost down," Clough asserts, but the development cost would still be greater than what is experienced for manned platforms. Using emergent behavior C2 concepts, however, presents the "possibility of removing entire blocks of code," he says. Clough sees the potential to reduce the lines of code to perform a task by 30 to 50 percent.

Emergent Software

The term, "emergent behavior," can be used to describe the process of developing software, as well. AFRL recently completed internal simulations of up to six UAVs, using trial-and-error methods to modify "generations" of algorithms as a way to identify the ones best suited to a UAV formation flight management task. AFRL contractor, Orbital Research, Cleveland, Ohio, is developing software tools "to breed emergent control behaviors," using genetic algorithms and programs. Evolutionary computational methods mimic natural selection by passing algorithms through hundreds or thousands of generations in a short period of time. Clough hopes to develop a software tool kit, or application, that can be reused by emergent systems researchers without having to be experts in the field. Orbital also will examine the problem of changing emergent behavior on the fly.

Seeking Weapons of Mass Destruction

Under another small business AFRL contract, Gild Associates in Columbus, Ohio, will build a distributed sensor system, based on UAVs, to detect agents for weapons of mass destruction, such as radioactive materials, bio-organisms and toxic chemicals. If the simulations in the first phase are successful, the company will test the concept with small UAVs.

An emergent "flight formation manager" algorithm could be available in two to three years, Clough says. But researchers would still have to convince a military flight safety board that it would be safe to try. Clough aims to remove any random activity in tasks like formation flight.

A Defense Advanced Research Projects Agency (DARPA) program, known as Software Enabled Control (SEC), also is considering algorithms for formation flight management of multiple UAVs but is deliberately avoiding the use of emergent behavior algorithms. According to a DARPA official, the program is exploring "group dynamic" algorithms, which "generate behaviors that are possible as a group but not possible as individuals."

To accomplish formation flight over rough terrain, for example, it makes more sense to give the group a single path, rather than provide each UAV with a detailed flight trajectory. "Such formations then can be maintained with locally generated maneuvers, using adaptive control systems," he says. SEC has found that passive sensing of neighboring vehicles’ positions and velocities enables the coordination of flight paths and can be effective in station keeping and achieving coordinated, collision-free group flight paths.

The DARPA program assumes the existence of external UAV command and control, in which the unmanned air vehicle "team"–not swarm–"must be responsive to commands and never do anything unexpected, outside the bounds of their intended maneuvering behaviors."

Long Pole in the Tent

How can developers make sure that swarms, or groups of UAVs using an embedded emergent behavior algorithm, do what they are intended to do and nothing else? How can they prove that nondeterministic software does what it is designed to do with an acceptable statistical probability?

While swarming technologies promise the eventual simplification of UAV control architectures, acceptance will be "the long pole in the tent," predicts Scheidt. It is very difficult to "program" a swarm to reliably do something that requires a coherent action in response to a user’s command, adds the DARPA official. "It has been found over the past few years that large swarms of entities, each acting on a reactive basis to local data, can be difficult to control for mission-level tasks." In the realm of air vehicles, attention needs to be paid to the potential unpredictability of their behavior, he stresses, since "air vehicles, more than ground vehicles, demand large safety and reliability margins."

The chances may be better for getting approval to use emergent behavior algorithms embedded within a more predictive C2 system. The AFRL’s Air Vehicles Directorate last year launched a research program with Lockheed Martin Aeronautical Systems, called Verification and Validation of Intelligent and Adaptive Control Systems (VVIACS). Implementation of new intelligent and adaptive flight control technologies is still some years away, but the program aims to have certification techniques ready by the time they are needed, says AFRL program manager Vincent Crum. While flight certification is the ultimate goal, it is beyond the program’s scope and budget.

Systems That Learn

"Existing [verification and validation] methodologies say all systems have to be deterministic, but human beings aren’t deterministic," Clough points out. With emergent behaviors one can’t expect that every input into a system can be traced to a specific output. "You’re talking about systems that learn, systems that negotiate," he says. Those are nondeterministic processes. "The systems are going to act differently from what you’d expect. But we need to make sure that that rate is kept commensurate with all the other failure rates of the system."

Verification and validation is the process of proving software’s operation and correctness. Verification usually refers to showing that the code fulfills its requirements, while validation refers to the completeness of the requirements. Unfortunately, "traditional V&V is expected to be completely ineffective for nondeterministic software," such as adaptive and intelligent control code, Crum explains.

VVIACS has conducted an industry survey that identified 48 research and development programs using emerging methods for future advanced flight control concepts. The program is analyzing the software structures and algorithms to help understand the problem, says Crum. After this phase, the program will begin developing methodologies for the verification and validation of advanced control software.

"Our ultimate goal is safety," says Crum. "The most commonly agreed-upon factor is one failure in 107 flight hours of operation." That number, however, is difficult to prove and in many cases is "based largely on statistical analysis of potential failure mechanisms."

Program officials have not decided how to prove safety. "But we fully intend to explore options beyond the traditional prediction analysis," Crum asserts. Although final certification of military air vehicles is done through the Defense Department, the program is seeking the involvement of the Federal Aviation Administration.

Avionics Magazine has reported on the U.S. Office of Naval Research’s (ONR’s) Autonomous Intelligent Networks and Systems (AINS) project (November 2002). AINS envisions a cooperative group of vehicles forming a robust, adaptable and scalable sensor and communications network, which can take a high-level human order and continue the missions even when lines of communication with battle commanders are broken.

Silver Fox

The AINS program sponsored exercises in August using experimental vehicles such as Silver Fox, a fixed-wing UAV built by Advanced Ceramics Research (ACR), in Tucson, Ariz. Courtesy of advanced AINS software built by institutions like the University of California at Los Angeles (UCLA) and UC Berkeley, Silver Fox took part in multiple networked missions. Though itself a prototype, the small unmanned air vehicle has been used in Iraq to fly convoy escort, reconnaissance and surveillance missions for Navy special forces and Marine Corps units.

The 8-foot (2.4-meter)-wingspan, model airplane-like UAV, originally developed for Navy whale tracking, flies for about five hours, with a 25-mile range, and can be equipped with medium-resolution color, high-resolution black-and-white and infrared cameras. There also are GPS for navigation and electronics for attitude and flight control. The Navy wants to increase endurance to 20 to 24 hours and shift from gasoline to diesel fuel. The Silver Fox’s maximum airspeed is 60 knots, although dash speeds in excess of 100 mph are claimed. Constructed of plastic, fiberglass and foam, the vehicle weighs 20 pounds (9.1 kg) fully loaded.

In one AINS exercise two Silver Foxes performed "auto-follow" maneuvers, flying alongside a bus. They were controlled by a pair of cell phones that relayed commands from the bus to a PC-based, mobile control station. Positioned on each side of the bus, the UAVs flew "sine wave" patterns, exchanging stations without colliding. In another demo people transmitted commands directly to the ground station via cell phones. Use of inexpensive communications devices could reduce the cost and increase the flexibility of UAV operations.

A car also was used to trigger ground sensors that formed part of a prototype AINS network. The ground sensors, detecting the car, called two Silver Foxes over to have a look, recalls Tony Mulligan, ACR’s chief executive officer. The sensors provided an approximate GPS location, and an onboard UAV visual system identified the object as a car. Silver Fox then called in ground robots, which surrounded the car–all without human intervention.

The demos kicked off a new round of AINS small business technology transfer contracts. In one effort, ACR will demonstrate search-and-rescue, border patrol and over-the-hill reconnaissance using fixed-wing and rotary-wing UAVs–the latter supplied by Massachusetts Institute of Technology (MIT). The UAVs will fly in formation with a manned helicopter. When the manned helicopter hovers, the four UAVs will orbit around their positions alongside the helicopter, extending its field of view.

Another ACR task under the AINS program will be to demonstrate convoy protection. Silver Foxes and perhaps one or two unmanned helicopters would identify and track targets, follow vehicles and provide protection against ambush.

The Navy also wants to apply the auto-follow technology to small, fast boats, where the boat’s direction controls the aircraft. Planned for the fall is an experiment to launch the vehicle from small patrol boats, according to Capt. John Hobday of ONR.

In addition, the UAV has demonstrated the dropping of spherical sensors attached to its wings. These electronically guided, sometimes winged, "eggs"–with glide slopes ranging from 2-to-1 to 15-to-1–have provided video data.

Bomb Damage Assessment

Truly lightweight UAVs, known as micro air vehicles (MAVs) may also find a battlefield niche. The Air Force Research Laboratory’s Munitions Directorate at Eglin AFB, Fla., plans a demonstration in about two years to test their use in bomb damage assessment (BDA).

A 24-inch (61-cm)-wingspan, 10.5- to 14-ounce (300- to 400-gram) prototype, constructed by AFRL of "pretty indestructible" carbon fiber material, would be released at a selected altitude from a guided bomb. As the bomb impacts, the MAV, powered by an electric motor, would orbit around preselected coodinates, transmitting images to a command facility.

This "instant BDA" would include color or monochrome imagery, showing the impact area, says Air Force 2nd Lt. Oluyomi Faminu, AFRL’s program manager for BDA. Avionics include cameras, GPS, an autopilot and flight control electronics. The goal is for the vehicle to operate autonomously, with the option to be handed off to a pilot, but initial tests will be remotely controlled via line-of-sight link from the ground.

The UAV’s electric motor produces speeds of 30 knots, so tight wind constraints are necessary. And it must be tough enough to survive the shock wave of the explosion. Flight duration will be about 30 minutes. A smaller version, less than 3.5 ounces (100 grams), will be a target BDA platform if the initial vehicle proves successful. It would require miniaturization of the sensors, autopilot and power source.

Next-Generation MEMS

As unmanned air vehicles get smaller, so will avionics gear such as inertial measurement units (IMUs). BAE Systems’ first-generation, solid state, silicon IMU (SiIMU), built using micro electromechanical systems (MEMS) technology, has been selected for South Africa’s Unmanned Aerial Observation System (UAOS) tactical platform, a 275-pound (125-kg) multimission UAV with a maximum altitude of 16,400 feet.

A second-generation BAE Systems IMU already is nearing availability. Expected to enter production in 18 months, the unit has been designed into the U.S. Army’s Advanced Precision Kill Weapon System, a 2.75-inch (7-cm) diameter guided rocket, and other programs. The second-generation design is specified for drift of less than 5 degrees per hour, once the initial turn-on error has been removed. That compares with 100 degrees per hour for the first-generation product.

Electronics consolidation is expected to reduce IMU size from 8 to 3 cubic inches (131 to 49.2 cubic cm). Weight likewise is specified at about 4 ounces (112 grams), versus 8.75 ounces (250 grams). The second-generation device will replace analog application specific integrated circuits (ASICs) with a digital design in which a single microprocessor will provide all the electronics functions. "Digital design gives you performance at very high spin rates" without IMU saturation, asserts a company official. In missile applications, for example, "We can adjust the measurement range of the IMU in real time, during the mission," to match the spin rate experienced at different phases of flight.

Receive the latest avionics news right to your inbox