This paper is the first in a two-part series that introduces an
easy-to-implement central command architecture for high-order autonomous
unmanned aerial systems. This paper discusses the development and the second paper presents the flight test results. As
shown in this paper, the central command architecture consists of a central command block, an
autonomous planning block, and an autonomous flight controls block. The central
command block includes a staging process that converts an objective into tasks
independent of the vehicle (agent). The autonomous planning block contains a
non-iterative sequence of algorithms that govern routing, vehicle assignment,
and deconfliction. The autonomous flight controls block employs modern controls
principles, dividing the control input into a guidance part and a regulation part.
A novel feature of high-order central command, as this paper shows, is the
elimination of operator-directed vehicle tasking and the manner in which
deconfliction is treated. A detailed example illustrates different features of
the architecture.
Unmanned aerial vehicles (UAV) are gaining interest with relaxing restrictions in civilian airspaces. Looking ahead, a systematic approach is needed for autonomous unmanned aerial systems (AUAS) consisting of a large number of vehicles (agents). This paper, which is the first in a series of two papers, presents an easy-to-imple- ment architecture for high-order AUAS; the second paper presents flight test results. The different AUAS ap- proaches are referred to in this paper as command approaches with central command at one extreme and indi- vidual command at the other. In central command, the AUAS authority lies in one entity, and the vehicle re- quires only limited information about its environment. Central command is naturally suited to problems dictated by global objectives, such as search and surveillance in precision agriculture, border security, law enforcement, and wildlife and forestry services, to name a few. In individual command, the authority lies in the individual ve- hicle. The vehicle requires local situational awareness, communication with other vehicles, and understanding of system objectives. Individual command is naturally suited to problems dictated by local objectives that require coordination, like those found in air traffic control problems and in coordinated pick and place problems.
When an AUAS operates multiple vehicles in close proximity some method of preventing collisions becomes necessary. Collision avoidance, or the real-time act of recognizing and maneuvering to avoid an impending con- flict, also called “see and avoid”, is a fundamental principle of current flight operations and will be a require- ment of future operations that share manned airspace. In contrast, deconfliction, or the act of planning paths in advance that do not conflict, is a central command level method of preventing conflicts as events are planned. Collision avoidance and deconfliction are not mutually exclusive; deconfliction prevents collisions between planned or known events whereas collision avoidance prevents collisions between unplanned events. Future sys- tems will incorporate both methods. This paper develops a robust method of deconfliction based on an assign- ment method that produces non-crossing paths. This method is presented within an easy-to-implement architec- ture that is scalable to high-order AUAS. The method section introduces the basic architecture and the results section provides a detailed example.
The requirements of an architecture under which high-order AUAS become feasible is the subject of ongoing research. Bellingham, Tillerson, et al., 2003 [
The HOCC architecture is described by the block diagram shown in
Block diagram for the HOCC architecture
The complexity of the individual processes used, namely the Visibility Graph method, the A* method, the Hun- garian method, are well known. For example, the number of calculations in A* is on the order of nmE log(V) where n is number of vehicles, m is number of tasks, E is number of obstacle edges, and V is number of obstacle vertices and the number of calculations in the Hungarian method is on the order of n4. The elimination of exter- nal iterations, as described below, leads to a robust, computationally efficient algorithm. Throughout this paper, we only consider vehicles that are assignable to every task (homogeneity).
As shown in
The staging process converts an objective into a set of tasks. This paper considers spatial staging, such as passing over a point (for imagery or delivery), a line scan (following a line segment) which could be part of covering a wider region, and loitering (orbiting a point) which could arise when a vehicle is waiting for further instructions or to monitor a point of interest. The spatial tasks have starting points, so the conversion of an ob- jective into a set of tasks determines a set of spatial points that vehicles need to reach. During the staging process the candidate routes are unknown and the assignment of vehicles to tasks is unknown.
Autonomous planning is the second block of the HOCC architecture and routing is its first process. Support tasks, such as launching or recovering and refueling, are conducted by the administrator (A) who operates in the background. The routing process itself is performed in two parts, mapping the environment and choosing a path [
A frequent goal in AUAS path planning, because of limited fuel and flight duration requirements, is to deter- mine vehicle routes by minimizing distance of travel of individual vehicles. An important variation on this goal, which is introduced in this paper, is the minimization of total distance of travel of all of the vehicles (a min-sum assignment). The minimization of total distance of travel leads to an importing routing principle: The routes de- termined by minimizing total distance of travel do not cross. The mathematical proof of this routing principle is presented in the Appendix. Although non-crossing is guaranteed, the paths can still touch obstacles. Indeed, ve- hicle-task segments can share vertices of obstacles, as shown in
As shown, two vehicle task routes share vertex C. The vehicle at A travels to C and then to A*. The vehicle at B travels to C and then to B*. Notice that the total distance of travel after the shared vertex, CA* + CB*, is the same whether A travels to A* or to B*. Indeed, the solution to the minimization problem is indeterminate when two routes share a vertex. The presence of this indeterminacy, however, presents an opportunity. It allows the algorithm to ensure that the vehicles do not conflict after passing the shared vertex C, such as a task at A* re- quiring a vehicle to loiter for a short time, thereby staying very near the path of the vehicle proceeding to B*. Assume that the vehicles begin at the same time and that vehicle B reaches C before vehicle A. The algorithm selects the vehicle that first reaches vertex C to travel to the farthest point, in this case, the vehicle at B travels to B*. The inspection of a potential conflict associated with a separation distance falling below a minimal level is thus reduced to checking separation distances while the vehicles are traveling along the two segments AC and BC before the shared vertex. The desire to ensure that vehicle separation distances exceed required minimums makes total distance of travel an important cost function in AUAS path planning. Thus, the min-sum assignment enables rapid deconfliction of planned paths by checking only shared vertices.
The assignments that minimize the total distance of travel are determined in this paper by the Hungarian me- thod [
Shared vertex
or vehicles.
As described above and proven in the Appendix, potential conflicts in this HOCC architecture are treated mainly by avoiding them in the first place (in contrast with collision avoidance). As stated earlier, this was accom- plished by utilizing the non-crossing property of a min-sum assignment. Non-crossing is a more powerful state- ment than deconfliction. A deconflicted route means two vehicles must not occupy the same space at the same time, but the non-crossing property shows that the entire length of the paths do not cross at any time. This de- couples vehicle speed from the path planning and is useful in solving the potential conflicts in the vicinity of ob- stacle vertices. By loosening any initial assumptions about vehicle speed, shared vertices can be deconflicted by small speed adjustments without changing the path of the vehicle. Because the assigned path is known not to in- tersect with other paths, this small speed adjustment cannot cause other routing conflicts, eliminating the need to re-check the final solution for conflicts The elimination of route intersections leaves potential conflicts only in the neighborhood of boundaries of obstacles that are shared by more than one vehicle. When a potential conflict is identified, the first step is to determine whether or not a conflict arises. The simplest action is to eliminate a conflict by adjusting the speed of one or more vehicles without changing any routes. Minor speed adjustments around the cruise speed are almost always feasible. In extreme cases timing can be corrected by placing vehicles in holding patterns, and is scalable to high-order AUAS when route segments are sufficiently long. Other reme- dies, like route adjustment [
The autonomous planning block leads to vehicle assignments. The vehicle assignments represent low frequency and bandwidth information that can be performed off-board. The autonomous flight controls, in contrast, re- present higher frequency and bandwidth processes that are ideally suited for on-board implementation. Best practices in autonomous flight controls are reviewed below. The first step is to convert assigned routes to guid- ance paths (G). At this stage routes are refined to account for turning radii and other properties to enable ve- hicles to follow the routes precisely. Indeed, slopes and curvatures of connected line segments are discontinuous at route vertices and can’t be followed precisely. Circular arcs between line segments [
guidance state
vehicle, the autonomous navigator, the autonomous regulator and the form of the control input are
where KGk is the guidance operator, KRk is the regulator operator, and where the control input is the sum of a guidance input uGk, a regulator input uRk, and a disturbance input uDk. This is called the modern control form [
where I is the identity operator and ek is the state error vector. The characteristic equation is independent of the guidance state vector. Therefore, the regulator gain matrix GRk determines the vehicle’s stability characteristics (settling time, peak-overshoot, and steady-state error) independent of the guidance state vector rk. These best practices pertaining to autonomous flight controls are summarized as an AUAS Flight Controls Principle: Assume that a realizable guidance path has been produced and that a vehicle has the control authority capable of maintaining the guidance path within certain limits placed on settling time, peak overshoot, and steady-state error. Then autonomous guidance and autonomous regulation can be designed by separate and distinct methods. Following best practices in autonomous flight controls lead to vehicle paths that more closely follow their gui- dance paths, which is of particular interest in autonomous systems that are not instrumented with on-board collision avoidance systems.
Four vehicles loiter in an air space that contains a no-fly zone (obstacle), as shown in
The scenario begins with the AUAS user commanding the first objective to survey area A. The command to survey area B is made later. The staging process converts the objective to search area A into 5 line segment tasks after which the vehicles are tasked to return to their original loiter points (see
Problem statement
Part of the graph between a vehicle and the end point of a task
After determining the possible graphs between a vehicle point and a task, the next step is to find the shortest route between them. The shortest vehicle-task routes between one of the vehicles and each of the five tasks are shown in
Once the shortest vehicle-task routes are determined, vehicles need to be assigned to tasks. This is done by minimizing the total distance of travel of the vehicles (see
In
Five shortest vehicle-task paths for one of the ve- hicles during the 1st time interval
Routes over the 1st time interval
A new objective during the 1st time interval initiates the 2nd time inter- val
Routes over 2nd time interval
process of executing a task are assigned to new tasks from the end point of the task currently being executed. When the new objective arrived, the blue and yellow vehicles were performing tasks, so their proposed paths beyond their current tasks were also included in the calculation. The proposed route of the yellow vehicle is to the left end point of the task just below its current task and the proposed route of the blue vehicle is to the right end of the task just above its current task. The red vehicle, after the new calculation, is assigned to the same end point as in the previous time interval.
The yellow vehicle is the first vehicle to complete a task. The completion of its task triggers the start of the third time interval.
In
In
In
In
Routes over the 3rd time interval
Routes over the 6th time interval
Routes over the 9th time interval
is routed to the farthest remaining vertex.
This paper presented an easy-to-implement central command architecture that is well suited to a large number of unmanned aerial vehicles—overcoming present limitations pertaining to operator task load, computational com- plexity and vehicle deconfliction. The architecture consists of a central command block, an autonomous plan- ning block, and an autonomous flight controls block. Building on results found in the literature pertaining to
Routes over the 10th time interval
12th (last) time interval
The end
AUAS, the central command block is fully autonomous thereby overcoming user load limitations associated with a large number of vehicles. The autonomous planning part is a non-iterative sequence of algorithms that govern routing, vehicle assignment, and deconfliction. The non-iterative nature of the algorithms overcomes computational limitations. The routing exploits a new routing principle that naturally prevents vehicle routes from crossing. An example illustrates the different features of the architecture. This paper is the first in a series of two papers; the second will report on flight test results.