Aero-Astro Magazine Highlight

RAVEN: Testbed for autonomous UAVs

By Jonathan How

Unmanned aerial vehicles have become vital warfare and homeland security platforms. They significantly reduce costs and the risk to human life, while amplifying warfighter and first-responder capabilities. These vehicles have been used in the Iraq war and during Hurricane Katrina rescue efforts with some success, but there remain technological barriers to achieving the vision of multiple UAVs operating cooperatively with other manned and unmanned vehicles in the national airspace and beyond.

Its four rotors frozen by the camera flash, a quadrotor hovers above Aero-Astro Professor Jon How’s hand. Flying completely autonomously, the UAV will stay within a 20cm box-shaped space — a feat exceeding the capabilities of the best human pilots. (William Litant photograph)

Key among the challenges to future unmanned vehicle use is the need to extend and distribute the team planning and control to achieve coordinated UAV behaviors in the presence of significant uncertainty about the operating environment. For example, consider a scenario in which a team of UAVs responds to a disaster site, maps the region for the human response team, searches for victims, interviews witnesses, and monitors the overall environment for threats. The UAVs must coordinate to ensure that the many tasks of this mission are performed as efficiently and effectively as possible, given the uncertainty in the terrain and poor knowledge of the scope of the mission. In another scenario, a team of UAVs could be used to protect a military convoy — fanning out to provide full surrounding surveillance while providing the ability to split off to take a more detailed look at targets of interest. Then, as fuel is used up, or unexpected mechanical problems occur, the UAVs would self-diagnose the problem, summon a replacement vehicle, and land on a mobile support platform to be replenished or sent to repair. In such a mission, the UAVs would coordinate to ensure that they are located in the best vantage points, given the terrain, to maintain full coverage around the convoy, and to ensure that routine, or unexpected, mechanical failures do not result in significant degradation in the team performance.

Researchers have recently developed algorithms to achieve cooperative UAV behaviors. A key step towards transitioning these high-level algorithms to future missions is to successfully demonstrate that they can be implemented in testbeds that use similar-sized (or scaled-down) vehicles operating in realistic environments. Doing experiments on scaled testbeds highlights the fundamental challenges associated with performing multi-day autonomous system operations with numerous human operators managing both high-level mission goals and autonomous UAVs conducting individual tasks. These challenges include planning for a large team in real-time with computation and communication limits; developing controllers that are robust to uncertainty in situational awareness, but are sufficiently flexible to respond to important changes; and using communication networks and distributed processing to develop integrated and cooperative plans.

Researchers have also developed a variety of research platforms to study advanced theories and approaches in the development of innovative UAV concepts. However, these testbeds typically have several limitations. For example, current outdoor platforms can be tested only during good weather and environmental conditions. Since most outdoor UAV test platforms can be flown safely only during daylight operations, these systems cannot be used to examine research questions related to long-duration missions, which may need to run overnight. In addition, many of these vehicles are modified to carry additional vehicle hardware for flight operations. Redesigned to meet payload, onboard sensing, power plant, and other requirements, they must be flown in specific environmental conditions, unrelated to flight hour constraints, to avoid damage to the vehicle hardware. These external UAVs also typically require a large support team, which makes long-term testing difficult and expensive.

The Aerospace Controls Laboratory
The MIT Aeronautics and Astronautics Department’s Aerospace Controls Laboratory set out to research and overcome these challenges. ACL investigates estimation and control systems for aerospace systems, with particular attention to distributed, multivehicle architectures. Example applications involve cooperating teams of UAVs or formation-flying spacecraft. The research goal is to increase the level of autonomy in these systems by incorporating higher-level decisions, such as vehicle-waypoint assignment and collision avoidance routing, into feedback control systems. Core competencies include optimal estimation and control, optimization for path-planning and operations research, receding-horizon/model predictive control, and Global Positioning Systems. This research has been demonstrated on several testbeds, including a team of eight rovers operated indoors and seven UAVs that are flown outside.

ACL experiments highlighted both the benefits and limitations of the testbeds. Based on these experiences, we developed the Real-time indoor Autonomous Vehicle test Environment, or RAVEN. RAVEN’s purpose is to examine long-duration missions in a controlled environment. The facility is designed to test and examine a wide variety of multivehicle missions using both autonomous ground and air vehicles. A key RAVEN feature is a global metrology system that yields accurate, high bandwidth position and attitude data for all vehicles in the room. Since the position markers are lightweight, the position system can sense position and attitude without adding significant payload to the vehicles. Thus the platform can use small, essentially unmodified, radio-controlled vehicle hardware such as electric helicopters and airplanes. This frees researchers from overly conservative flight testing, and allows us to simultaneously fly eight or more air vehicles in the confines of the lab.

An additional benefit is that one operator can set up the platform for flight testing multiple UAVs in fewer than 20 minutes, so researchers can perform a large number of test flights in a short period of time with little logistical overhead. Furthermore, since the system autonomously manages the navigation, control, and tasking of realistic air vehicles during multivehicle operations, researchers can focus on the algorithms associated with the team coordination rather than the details of the implementation. These properties greatly enhance the utility of the testbed, making it an effective platform rapid prototyping environment for multivehicle mission management algorithms.

Graduate student Per Adrian Alexander Frank observes a foam aircraft at it hovers autonomously nose-up in the Aerospace Controls Lab — an extremely challenging, and successful, test. (William Litant photograph)

Architecture
In our lab, the control algorithm and command data for each vehicle is processed by a dedicated computer and sent over a USB connection from the vehicle’s control computer to the trainer port interface on the vehicle’s transmitter. All computing for this system is performed on ground-based computers and a Vicon MX camera system measures the position and attitude for each vehicle in the testbed at rates up to 120 Hz. This motion capture system provides a simple, baseline capability for sensing and controlling the vehicle motion, which enables researchers to explore research topics, such as multivehicle coordination, vision-based navigation and control, or new propulsion mechanisms such as flapping flight.

Just as GPS spurred the development of large-scale UAVs, we expect this new sensing capability to have a significant impact on 3D indoor flight, which has historically been restricted to very small areas.

Results
RAVEN comprises a variety of rotary-wing, fixed-wing, and ground-based R/C vehicles. However, most testbed flight experiments are performed using the Draganflyer V Ti Pro quadrotor. While easier to fly than a standard helicopter, quadrotors are unstable and there is a strong coupling between the attitude control and the position loops. The four motors’ speeds must be rapidly and precisely adjusted to balance the vehicle and overcome any external disturbances. Thus quadrotors are very difficult to fly manually without significant operator training. In our typical 10-minute tests, we have achieved the ability for the Draganflyer to hover autonomously in a dimension half the size of the vehicle itself. The quadrotor stays inside a 20-cm box during the entire flight, far exceeding the capabilities of our best human pilot.

Tests have also been performed with a foam airplane. We first tackled a nose-up hover condition, which is a very challenging flight configuration for an aircraft because there is limited airflow over the rudder and elevator that must be used to control the vehicle position and these surfaces are partially blocked by the wing ailerons which are used to offset the motor torques, which are varied to control the aircraft altitude.
Again, tests confirm that the vehicle can be made to autonomously hover within a 20-cm box more than 63% of the time. As with the quadrotor, these results far exceed the capabilities of our best pilots, and were developed in fewer than two months.

While these are impressive results are for a single vehicle, a more important feature of RAVEN is that we can routinely operate five vehicles at the same time, and have flown as many as ten.

RAVEN also provides a superb learning experience for a number of students, who have been instrumental in its development. Currently, RAVEN forms the basis for seven Aero-Astro and Electrical Engineering and Computer Science graduate students’ thesis research. Five Aero-Astro and Electrical Engineering and Computer Science student, working under MIT’s Undergraduate Research Opportunities Program, complete the team.

Our main goals as we move forward are to demonstrate the use of health management tools to improve the performance of a team of UAVs performing persistent surveillance tasks. We plan to demonstrate this, in collaboration with Boeing Phantom Works, during the summer of 2007. RAVEN is also being used to perform flight demonstrations of cooperative planning and control concepts under development for the Air Force Office of Scientific Research. We are using funding from the MIT Aeronautics and Astronautics Department and the School of Engineering for several additional research projects, including human interaction with autonomous robots; coordination, control, and sensor fusion over intermittent communication networks; control of micro and nano air vehicles, including flapping flight; autonomous acrobatic aircraft; flight control using other sensors suites, such as vision, using reduced Vicon information. RAVEN is an impressive facility for multi-vehicle testing — we have only just started to explore its full capabilities.

Jonathan P. How is an associate professor in the Department of Aeronautics and Astronautics at the Massachusetts Institute of Technology. His research interests include using operations research tools, such as mixed-integer programming, to optimize the coordination and control of autonomous vehicles in dynamic uncertain environments. He was the recipient of the 2002 Institute of Navigation Burka Award, is the Raymond L. Bisplinghoff Fellow for Aero-Astro, is an AIAA Associate Fellow, and an IEEE senior member. He may be reached at jhow@mit.edu