FLC-based Landing Approach and Collision ...

50th AIAA Aerospace Sciences Meeting including the New Horizons Forum and Aerospace Exposition 09 - 12 January 2012, Nashville, Tennessee

AIAA 2012-0489

FLC-based Landing Approach and Collision Avoidance Path Planner for Multiple Aircraft and Runways

Downloaded by UNIVERSITY OF CINCINNATI on January 31, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2012-489

Joseph W. Stern* and Kelly Cohen† University of Cincinnati, Cincinnati, Ohio, 45221

The complexities of safely landing multiple aircraft on multiple runways within a controlled airspace pose many issues for human air traffic controllers, as well as in potential future situations wherein UAVs are introduced. Fuzzy logic control provides many benefits with properties such as universal approximation and exploitation of expert-provided heuristics and can aid in providing safe landing approach path planning and collision avoidance. This paper addresses an egocentric, autonomous method, wherein each aircraft applies fuzzy logic locally to avoid potential collision with other, landing aircraft in the airspace (as opposed to exocentric, collaborative control, wherein all aircraft are aware of all flight paths and base decisions on this knowledge). Given a two-dimensional airspace with multiple aircraft—each at constant airspeed, with a minimum turn radius, and neglecting maneuvering dynamics—and multiple runways, a fuzzy logic controller was designed and coupled with a simple, geometric path planner and supporting simulation code in Matlab.

Nomenclature FLC = Fuzzy Logic Controller UAV = Unmanned Air Vehicle (remote-piloted or fully autonomous) FIS = Fuzzy Inference System own ship = the aircraft being commanded other ship = the aircraft posing a potential collision threat FAA = Federal Aviation Administration In the context of the linguistic variables and rule base design: L = directional location of other ship relative to own ship A = direction of other ship relative to own ship SW, NE = linguistic variables for L and A e.g. south-west or north-east, respectively R = ratio of own ship velocity over other ship velocity F, M, S = linguistic variables for R: fast, medium, and slow, respectively D = normalized distance between own ship and other ship N, M, F = linguistic variables for D: near, medium, and far, respectively TA = commanded turn angle of own ship P, N = linguistic variable prefixes for TA: positive and negative, respectively S, M, L = linguistic variable suffixes for TA: small, medium, and large, respectively TR = commanded turn radius of own ship as a factor of its minimum turn radius T, M, L = linguistic variable suffixes for TR: tight, medium, and loose, respectively

I.

Introduction

A

IR traffic control is currently a very complex and delicate issue (Figure 1), and this currently involves only human-piloted aircraft. The future introduction of UAVs further complicates this. However, fuzzy logic could help alleviate some of the difficulty by introducing automation of landing approach and collision avoidance; even providing the air traffic controllers with only suggested course changes could prove beneficial. In addition, allowing * †

PhD student, Department of Aerospace Systems, Mail Stop 0070, student member. Associate Professor, Department of Aerospace Systems, Mail Stop 0070, AIAA member. 1 American Institute of Aeronautics and Astronautics

Copyright © 2012 by the American Institute of Aeronautics and Astronautics, Inc. All rights reserved.


UAVs to land themselves while avoiding all other aircraft in the airspace would prove useful especially if direct, remote piloting failed and an automated guidance system was necessitated.

Figure 1. Illustrative depiction from Ref. 1. An inspiration for this research came from the video game, Flight Control [4] (Figure 2), available on many platforms. The objective of this game is for the user to guide multiple aircraft to multiple runways without collision; each aircraft has a constant airspeed, many aircraft types are provided with different airspeeds and runway requirements (e.g. helicopters to the helipad, small single-prop aircraft to the short runway, and jet aircraft to the long runway), and all aircraft operate in a 2D playing field. The turning radii are minimal (airplanes can turn around as fast as the helicopters), but the principle objective of collision avoidance provided insight into the algorithm to be implemented in the FLC here.

Figure 2. Screenshot from Ref. 2.

II.

Design Background

Following previous FLC designs based on two-vehicle collision avoidance objectives3, Matlab’s Fuzzy Logic Toolbox was utilized to produce a FIS for directing a single aircraft (“own ship”) away from potential collision with another aircraft (“other ship”). The FIS design draws largely on researched collision avoidance work; overall principles of operation were guided by researched FIS-based obstacle avoidance work4. The FLC design also benefited from the capture of operator or artisan knowledge 5, which was developed through extensive, hands-on research2. The FIS has four inputs: distance (normalized by some factor), speed ratio (of the own ship speed over the other ship speed), bearing (direction from own ship to other ship), and heading (of other ship, relative to own ship). Distance has three membership functions describing “far,” “medium,” or “near” ranges. Speed ratio has three membership functions describing “fast” (own ship considerably faster than other ship), “medium” (both ships roughly the same speed), and “slow” (other ship considerably faster than own ship). Bearing and heading are both 2 American Institute of Aeronautics and Astronautics


divided into eight regions; the cardinal directions are slightly slimmer than the others, with naming conventions that reference the own ship’s frame. The FIS has two outputs: turn angle (left or right, along with a magnitude) and turn radius (normalized factor of the minimum turn radius). Turn angle has membership functions describing the magnitude of the desired turn (“far,” “medium,” and “small” deflections) in both the positive and negative directions. Note first that a “zero” angle was not defined. Also note that the FIS is evaluated continuously during flight, signifying that only turn direction is really necessary. Turn radius is a factor of the own ship’s minimum turn radius, ranging from one to four times this value. Additionally, limited tuning was performed for this initial concept evaluation, resulting in the output hypersurface for turn radius having a minimum value of 1.5 and not 1; for simplicity, this output was divided by 1.5 in order to allow minimum-radius turns to be achieved when needed. As avoiding collisions through more evasive, tighter turns is a high priority, reducing the maximum output of this hypersurface was of minimal concern. The FIS resulted in 22 input and 10 output membership functions and was then designed with 132 rules (samples shown in Figure 3), covering the spectrum of approach angles and directions, as well as speed and distance combinations. If the normalized distance was greater than unity, then there was no collision threat to consider and the FIS was not evaluated for that aircraft. L SW

A N

R F

M

S

NE

NW

F

M

S

D N M F N M F N M F N M F N M F N M F

TA PS PS NM NM NS PL PL PL PL PL PL

TR M L T M L T T M T T M

Figure 3. Sample rules for FIS.

III.

Simulation Setup

The FIS is only a small kernel of the entire system. In order to arrive at the FIS, all aircraft positions, velocities, and normalizing values for distance need to be computed at each selected time step in the simulation. In order to apply the results of the FIS to the vehicle’s flight, a path planner was written in Matlab. This was a simple, geometric path planner: two constant-radius arcs to turn the aircraft initially onto the transit path and then finally onto the approach path and the straight-line segment in between were computed. Some robustness was applied for situations where the aircraft was commanded to deflect from its approach path about a radius that does not immediately realign it with the final approach segment. These components—the input generator, the FIS, and the path planner—were integrated into a Matlab simulator which plots all results in either of two manners. If the airspace is relatively simple and decluttered, the final results are immediately displayed for evaluation and tuning of the FIS. If the airspace is too dense and cluttered, then the simulation plot steps through each time step, thus animating the behavior of the aircraft. With a working simulator, several tests were progressively performed; for all tests, the aircraft operated on a 2D plane and maintained constant airspeed and clean turns (simplified dynamics). The first and most simple involved two aircraft: the other ship was (unrealistically) given a constant airspeed and location, while the own ship was allowed to fly its landing approach and collision approach path. Since this initial test is so unrealistic, only one collision threat was considered; subsequent tests offer multiple collision threats. The next test involved the other ship having a constant airspeed and initial location, while the own ship maneuvered to avoid it. This was also unrealistic, as the other ship simply flew through the airspace and never landed, often resulting in the own ship (lacking higher-level logic to satisfy other objectives) never reaching the runway in a timely fashion. The next test involved multiple aircraft each automated to land (without any collision consideration) with the own ship commanded to land while actively avoiding collision. This required additional logic, in order to evaluate 3 American Institute of Aeronautics and Astronautics

the end effect of all FIS-based commands evaluated for each other ship and the own ship; since the turn angle output implicitly contains both the turn direction and a magnitude of turn which signifies the collision risk (i.e. the more imperative the maneuver results in a greater deflection), the FIS result with the greatest commanded deflection dictated the maneuver at that time step. Preliminary results are presented for this test. The final test involved allowing all aircraft to employ their own FIS and path planner, allowing full autonomy in the airspace. This way, multiple runways and aircraft configurations can be provided; a range of flight characteristics (typical airspeeds and minimum turn radii) was developed and used for all cases. In order to visualize the performance in complex scenarios, the airspace picture was animated in Matlab; the figure was updated at a regular rate in order to show temporally the dynamic interactions of multiple UAVs.


IV.

Results and Discussion

The results for the above simulations are shown below, accompanied by a discussion of the results. Most of the subsequent figures show the UAV of concern (the “own ship”) with blue markers, while the collision threat (the “other ship”) appears in red; the runways are represented with a green cross. The initial landing approach path for the own ship is shown. Each step in the progression of the simulation is shown; a breadcrumb is plotted as a black dot for the own ship or a red dot for the other ship. After 10 steps, a significant marker is shown as a blue or red arrow (to show UAV heading at that point) or blue diamond or red triangle. A. Test 1: Single Collision Threat, Constant Location and Heading Figure 4 shows one of the cases wherein the presence of a potential collision threat caused the FIS to divert the UAV from its initial course. The FIS evaluation suggested a constant-radius turn away from the other ship, resulting in the actual path flown. Case 3 5000

4500

4000

3500

3000

2500

2000 500

threat

1000

1500

2000

2500

3000

3500

4000

Figure 4. Single, constant-location threat; simple path diversion. Figure 5 shows one of the cases wherein the other ship is closer to the own ship on the side to which the previous case showed diversion. The FIS caused the own ship to deviate in the initial direction but at less of a turn radius than the initial path planner chose systematically. Due to the proximity of the other ship to the runway and the potential of collision during this new path, the own ship continues to deviate away from the runway until it is close enough to decide that it can land safely before the other ship can pose any more collision threat. Since the other ship is stationary with a constant speed, this result does not present a realistic analog; however, it clearly demonstrates the effectiveness of the FIS decision-making. The determined turning radii are evident in the figure as well.

4 American Institute of Aeronautics and Astronautics

Case 4 5000

4500

4000


3500

3000

threat

2500

2000 0

500

1000

1500

2000

2500

3000

Figure 5. Single, constant-location threat; multiple path diversions. Figure 6 demonstrates a scenario wherein the other ship’s proximity to the runway and to the own ship is just sufficient to cause the FIS to divert the own ship around the back side of the runway; this scenario required the simple, geometric path planner to be made robust enough to handle cases wherein successive deviations were generated within close proximity to the runway (i.e. the radius about which to turn the aircraft from its current location and the radius about which to steer it onto the final approach cause overlapping arcs). Case 5

5500

5000 4500

4000 3500

3000 2500

2000

0

500

1000

1500

threat 2000 2500

3000

3500

4000

Figure 6. Single, constant-location threat; advanced path diversion.


4500

Figure 7 demonstrates a scenario wherein the own ship is diverted into a reentrant path; the unrealistic flight condition of the other ship causes the current FIS to deviate so far from the landing approach so far away from the other ship course that it ends up completing a full turn and enters a new turn. While this specific scenario is unrealistic, this still exposes the potential for a UAV to fail to land in a timely manner, given the current FLC design. Case 6 5000

4500


4000

threat

3500

3000

2500

2000

-1000

-500

0

500

1000

1500

2000

2500

3000

Figure 7. Single, constant-location threat; reentrant path diversions. B. Test 2: Single or Multiple Collision Threats, Initial Location, Constant Heading Figure 8 demonstrates what happens when the other ship effectively blocks the own ship from successfully approaching the runway; due to closeness in aircraft speeds (i.e. the ratio input for the FIS is close to unity), the own ship spends a significant amount of time (and therefore fuel) trying to overcome the other ship. Since the FIS has no memory and provides no path optimization, these are the results that the UAV would currently face if it employed the designed FLC. However, as with the first test, this test is unrealistic in that some aircraft will fly completely across the controlled airspace while making no landing approach. Future FLC-based designs should account for this condition, though, and provide an alternative flight path.


Case 3

12000

10000


8000

6000

4000

2000

-2000

0

2000

4000

6000

8000

10000

Figure 8. Single, constant-heading threat with initial location; failure to overcome threat with excessive path diversions. Figure 9 shows the case wherein the own ship is significantly slower than the other ship; it is able to turn back to a landing approach path sooner, once the other ship has cleared the imminent airspace. This case also shows the difference in rules wherein the own ship is somewhere relatively behind the other ship; its flight path is less restricted, due to the significant decrease in collision potential. Case 7 10000 9000 8000 7000 6000 5000 4000 3000 2000

-1000

0

1000 2000 3000 4000 5000 6000 7000 8000

Figure 9. Single, constant-heading threat with initial location; ability to overcome threat with fewer path diversions. 7 American Institute of Aeronautics and Astronautics

Figure 10 shows the specific case wherein the own ship is able to deviate immediately to a path that places it behind the other ship, given the nominal conditions. Case 8 5000

4500

4000


3500

3000

2500

2000

1500 -1500

-1000

-500

0

500

1000

1500

2000

2500

3000

Figure 10. Single, constant-heading threat with initial location; own ship’s immediate path diversion to rear of other ship’s flight path. Figure 11 demonstrates a scenario with multiple collision threats; two other ships are provided, one on either side of the airspace along the own ship’s initial flight path. As can be seen, the FIS decided to deviate the own ship away from the more proximal threat and toward the more distant threat. As also can be seen, the current tuning of the FIS caused the own ship to deviate away from a nominal landing approach; as the evaluated FIS decided that the own ship was approaching the other ship on the left, the own ship was deviated toward the other ship on the right. This outcome may be due to the lack of symmetry in the rule base; guided by the maritime collision avoidance reference3, certain maneuver conventions were developed and implemented in the FIS. Head-on collisions are avoided by diverting the UAV to the right; the UAVs would separate and open the distance between them. Also, there is a maneuver to allow one UAV to overcome another; the slower UAV in front will divert to the left, while the faster UAV behind will divert to the right. The influence of these integrated maneuvers can be seen in the case.


Case 9 10000 9000 8000 7000


6000 5000 4000 3000 2000

-3000 -2000 -1000

0

1000

2000

3000

4000

5000

6000

Figure 11. Multiple, constant-heading threats with initial locations; path diversions to avoid collisions. C. Test 3: Single Collision Threat on an Uninterrupted Landing Approach Finally, the third test began to show some realistic results; with the other ship flying on an uninterruptible landing approach path, the own ship’s FIS has to determine how to land safely in a simplified but still realistically dynamic environment. Figure 12 shows a case wherein the own ship cannot land safely before the other ship. The FIS diverts the own ship into a sort of “holding pattern,” allowing the other ship to land safely before continuing. This also revisits the earlier case, which resulted in the FIS maintaining an excessive series of path diversions.


Case 11 5000

4500

4000


3500

3000

2500

2000

1000

1500

2000

2500

3000

3500

4000

4500

Figure 12. Single threat on an uninterruptible landing approach; path diversion into a small holding pattern to avoid collision. Figure 13 shows a case wherein the own ship is allowed to land safely before the other ship can approach. While this seems a very simple case, if reinforces the possibility that the UAV might not have to perform any complicated maneuvers and run the risk of diverting excessively, threatening the UAV through unnecessary flight time and fuel expenditure. Case 12 5000

4500

4000

3500

3000

2500

2000

1000

1500

2000

2500

3000

3500

4000

Figure 13. Single threat on an uninterruptible landing approach; no path diversion was deemed necessary in order to avoid collision. 10 American Institute of Aeronautics and Astronautics

D. Test 4: FLC Implemented on Multiple UAVs The final test demonstrates the dynamics of multiple FLCs diverting multiple UAVs in order to provide landing approach path planning along with collision avoidance. Figure 14 shows a case wherein two UAVs are attempting to land. One UAV is evidently slower than the other, given the spacing of the “breadcrumb” dots at constant time intervals. It is shown that the slower UAV allows the faster UAV to land first after they both begin diversions at about the same time. This is the more economical outcome, since the faster UAV would have a larger minimum turning radius and would cost more in order to wait long enough for the slower UAV to finally land. It can also be seen that the output for turn radius was actually a larger real value than the initial landing path; this was due to the arbitrary selection of an initial path at minimum turn radius; the FIS output did not necessitate such a tight turn in order to avoid collision. Case 15


5000

4500

4000

3500

3000

2500

2000

0

500

1000

1500

2000

2500

3000

Figure 14. Two UAVs attempting to approach the runway while implementing FLCs for collision avoidance; the slower allows the faster to land first. Figure 15 demonstrates a scenario wherein two runways exist for the two UAVs; furthermore, the selection of runways necessitates a crossing of initial flight paths. In this case, the two UAVs both begin diversions away from each other at about the same time; the faster UAV is forced to miss its nominal landing approach and must continue away from the collision threat region. During this time, the slower UAV’s FIS determines that the faster UAV has flown out of potential collision range and thus begins to allow the slower UAV back onto its landing approach path. The faster UAV is soon without any potential collision in the airspace and plans a landing approach path, thus exiting the large deviation behind the runways.


Case 16 6000 5500 5000 4500 4000


3500 3000 2500 2000 0

1000

2000

3000

4000

5000

Figure 15. Two UAVs implementing collision avoidance while attempting to approach two runways; the faster performs a large maneuver while the slower lands first. Figure 16 demonstrates the potential complexity that can be faced when several UAVs are all attempting to land within a confined airspace; whenever they have all diverted away from landing and have decided to return to a path to approach the runways, they also approach each other and then begin a new set of diversions to avoid collision. Figure 17 shows the initial setup for this case: four UAVs start off with flight plans to three runways (one runway is shared between two UAVs). This condition seems less similar to the case shown in Figure 14, wherein the UAVs approached the same runway, and more similar to what was seen in Figure 15, wherein separate runways are being approached by multiple aircraft. As can be shown in animated descriptions (more so than in the static, final picture), it appears that the slower UAVs generally are able to land first; by the time that the faster UAVs have overshot the runway region and can begin returning for approach, the slower UAVs have returned to flight paths not too different from their paths before the collision avoidance maneuver. Figure 18 more clearly shows the final results: the breadcrumbs were removed, solid and thin lines were generated between each point, and the division markers at constant time intervals remain clearly visible. This case especially demonstrates how the FLC provides universal approximation but not an optimal solution. More on this will be discussed in the following section.


Case 17 4500 4000 3500 3000 2500


2000 1500 1000 500 500

1000 1500 2000 2500 3000 3500 4000 4500 5000 5500

Figure 16. Four UAVs approaching three runways while implementing FLCs for collision avoidance. Case 17

3500 3000 2500 2000 1500 1000 500 0 500

1000

1500

2000

2500

3000

3500

4000

4500

5000

5500

Figure 17. Four UAVs approaching three runways while implementing FLCs for collision avoidance.


Case 17 5000 4500 4000 3500 3000 2500


2000 1500 1000 500 0 -1000

0

1000

2000

3000

4000

5000

Figure 18. Four UAVs approaching three runways while implementing FLCs for collision avoidance.

V.

Future Work

As this was only an initial concept study, limited tuning of the FIS was provided. As could be seen, issues in the performance of the collision avoidance FIS should be explored in detail. Aircraft flight dynamics should be considered more in-depth. As with the path planning algorithm employed in the obstacle avoidance reference4, sufficient “padding” was placed around the obstacles (in this case, dynamic obstacles) in order to avoid collision with safe clearance. The 132 rules were carefully chosen to reduce the required tuning of the 22 input and 10 output membership functions. Future work should explore Genetic Algorithm and Neural methods to automate tuning of the FIS. Advanced testing beyond a small collection of arbitrary cases should be performed; Monte Carlo methods should be employed to test a wide range of initial conditions for more than four aircraft with multiple runways. In addition, cascading fuzzy inference systems should be developed; currently, additional logic was developed to account for multiple collision threats, since the FIS makes decisions based on only one threat. Being able to weigh the threat of collision in a complex airspace using linguistic variables might aid in making decisions when there is very little room to maneuver, thus improving information equity5. The method presented in this paper was an egocentric, fully autonomous approach: each UAV’s FIS based decisions on the current state of all other aircraft. An exocentric, collaborative approach, wherein each UAV is aware of all other UAV’s flight path, should be considered. This would introduce linguistic variables concerning “path directness” instead of simple deviations and would allow flight path optimization for all UAVs. This would result in a contrasting report and also provide for a comparison of the two methods; possibly, an effort could be made to combine both methods, balancing path-consideration complexities with immediate-consideration-only simplicity to provide an optimal solution. Also, for the conditions where many UAVs have clustered into too dense of a region, coupling this method with a cascading FIS could carefully alleviate this situation or avoid it altogether. It was also suggested to consider airspace partitioning; it is possible that various aircraft types are restricted to confined regions, imposing additional constraints on the performance of the FLC. Again, cascading FIS design could help address this. There may be other constraints imposed by the FAA, which would require additional fuzzy logic.



VI.

Conclusion

An FLC-based algorithm to provide landing approach and collision avoidance path planning was presented. Several test cases were devised to determine the FLC’s performance. In fact, it was seen that the FLC itself was only a small kernel of the complete system; several components, including a simple path planner and additional logic to handle the advanced cases, had to be developed as well. The behavior of the FIS could be seen in the corresponding figures, showing the intended maneuvers and multiple-UAV dynamics throughout the scenarios. The final, realistic test revealed issues concerning the nature of the designed FIS itself: it is a universal approximator but does not provide an optimal solution. This was most evident in the final test case presented, wherein four UAVs attempted to land on three runways and had to perform many maneuvers in order to avoid collision. On occasion, the slower UAVs could deviate from their path to a runway shared with faster UAVs, allowing the faster UAVs to land first before the slower UAVs could easily return to their landing approach path. More often, the slower UAVs typically deviated slightly on approach to their runway while the faster UAVs deviated greatly (due to their speed) from their landing approach paths in order to avoid collisions; the faster UAVs performed large-radius “holding patterns” while the slower UAVs typically were able to land first. Several suggestions for future work were also presented. While this paper focused on an egocentric, fully autonomous method, an exocentric, collaborative method should be explored. This would provide an alternative approach to solve the same problem with a different solution and would also provide an opportunity to perform a comparison of the two methods and possibly combine them to balance the advantages and disadvantages of them both.

Acknowledgments The primary author of this paper would like to thank the primary author of the main reference 4 for direct guidance and support on the work done previously and also the secondary author of this paper for exposure to possibly the best way2 to develop an expert knowledge base for FIS design.

References 1

http://reason.com/blog/2008/08/28/us-air-traffic-control-the-mor http://firemint.com/flight-control/ 3 Perera, L. P., Carvalho, J. P., Guedes Soares, C., “Fuzzy logic based decision making system for collision avoidance of ocean navigation under critical collision conditions,” J Mar Sci Technol, 16:84–99, 2011. 4 Sabo, C., Cohen, K., Abdallah, S., “Effectiveness of 2D Path Planning in Real Time using Fuzzy Logic,” 48th AIAA Aerospace Sciences Meeting, 2010. 5 Thomas, D., Armstrong-Helouvry, B., “Fuzzy Logic Control – A Taxonomy of Demonstrated Benefits,” Proceedings of the IEEE, Vol. 83, No. 3, 1995. 2
