This document evaluates hardware platforms and path replanning strategies for emergency landing of unmanned aerial vehicles (UAVs). It compares the performance of greedy heuristic, genetic algorithm, and ensemble methods on personal computer and embedded computer hardware when responding to critical situations like motor or battery failure. The genetic algorithm and ensemble genetic algorithm-genetic algorithm method achieved the best results overall, finding near optimal emergency routes. Testing was done in simulation and with a real UAV to validate the approaches.

1. Evaluating Hardware Platforms and Path
Re-Planning Strategies for the UAV Emergency
Landing Problem
Jesimar S. Arantes, Márcio S. Arantes, André B. Missaglia,
Eduardo V. Simões and Claudio F. M. Toledo
November – 2017
Jesimar S. Arantes, Márcio S. Arantes, André B. Missaglia, Eduardo V. Simões and Claudio F. M. Toledo (USP)ICTAI 2017 November – 2017 1 / 22

2. Outline
1 Introduction
2 Problem Description
3 Methods
4 Computational Results
5 Conclusions
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 2 / 22

3. Introduction
This work addresses
Path replanning to land the UAV under critical situation:
Greedy Heuristic (GH) (Arantes et al. 2015).
Genetic Algorithm (GA) (Arantes et al. 2015).
Ensemble GA-GH.
Ensemble GA-GA.
Critical situations considered:
Motor failure.
Battery failure.
Supervision of safety systems
IFA: In-Flight Awareness (Mattei et al. 2013).
Performance evaluation:
Personal computer: Intel i5
Embedded computer: Intel Edison
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 3 / 22

4. Problem Description
dc
critical
situation
no-fly region
penalty region
bonus region
Legend:
mission route
emergencial route
aborted route critical distance
Figure 1: Illustrative scenario of emergency path re-planning.
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 4 / 22

5. Problem Description
Problem: Path-replanning problem
Minimize Pr
j
xT ∈ Op
j − Pr
i
xT ∈ Ob
j (1)
subject to:
xt+1 = FΨ(xt, ut) + ωt ∀(t) (2)
x0 ∼ N(ˆx0, Σx0 ), ωt ∼ N(0, Σωt ) ∀(t) (3)
Pr


j t
xt /∈ On
j

 ≥ 1 − ∆ (4)
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 5 / 22

6. Problem Description
Legend:
map boundary
mission routeno-fly region boundary
penalty region boundary
bonus region boundary takeoff
landing
Figure 2: Real world case study for aerial imaging application.
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 6 / 22

7. Methods
Compass
Sensors
GPS
Processor
AutoPilot
IFA System
Data
Acquisition
Sensor
Fusion
Security
Control
Decision
Making
Communication
Control
Path
Replanning
MOSA System
Data
Acquisition
Sensor
Fusion
Mission
Control
Camera
Control
Communication
Control
Path
Planning
embedded
embedded
Barometer
Accelerometer
Gyroscope
Follow Waypoint
Power Module
Power
Battery
Camera
Legend
Power supply
HW communication
SW communication
Actuators
Brushless Motor
Servo Motor
UAV
Modem Telemetry
Communication
Radio receiver
Methods
IntelEdison-DualCore
Figure 3: Embedded system architecture in the aircraft.
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 7 / 22

8. Methods
Path
Replanning
Map
Initial
State
Input
Dynamic
Model
Output
Config
Landing
Site
Fitness
Set of
Waypoints
IFASystem
Methods
Failure
Probability
of Landing
GA
GH
GA-GA
GA-GH
Figure 4: Embedded system architecture in the path re-planning module.
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 8 / 22

9. Methods
GA
evaluation
best route
failure
differents
methods
GA-GH
(a)
core1
GH
core2
GA
evaluation
best route
failure
same
methods
GA-GA
(b)
core1
GA
core2
Figure 5: Implemented strategies running methods in parallel. (a) Combining of
GA e GH. (b) Two executions of GA.
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 9 / 22

10. Methods
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 10 / 22

11. Methods
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 11 / 22

12. Computational Results
Settings used in the GA method.
Parameter Value Parameter Value
population size 39 crossover rate 0.5
tournament size 3 mutation rate 0.7
stopping criterion time elitism yes
Computer settings used.
PC i5 Intel Edison
Frequency 1.8 GHz 500 MHz
Memory RAM 4 GB 1 GB
Operating System Linux - Ubuntu Linux - Yocto
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 12 / 22

13. Computational Results
Technical specifications of the Ararinha.
Component Value Component Value
Wingspan 1.90m Weight 2.83kg
Length 1.15m Payload 0.60kg
Electric Power 740W Endurance 15 minutes
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 13 / 22

14. Computational Results
Time analysis for convergence of the GA.
200
260
320
380
440
500
560
620
680
740
800
860
920
980
1040
1100
1160
1220
1280
1340
1400
1460
1520
1580
1640
1700
1760
1820
1880
1940
2000
0
200
400
600
800
1000
1200
1400
1600
1800
2000
avg global
Time (ms)
Fitness
ε < 1%time = 890
ε < 2%time = 690
ε < 4%time = 460
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 14 / 22

15. Computational Results
GH
We run 1 time over each map
We evaluated 30 artificial maps in total
We evaluated 2 critical situations (motor and battery)
GA, GA-GH, GA-GA
We run 10 times over each map
We evaluated 30 artificial maps in total
We evaluated 2 critical situations (motor and battery)
We evaluated 3 stopping criterion based on time
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 15 / 22

16. Computational Results
n1
n2
p1
p2
p4
p3
b6
b7
b2
b3
b1 b4
b5
b8
(a)
n1
n2
p6
p7
p2
p3p1
p4
p5
p8
n4
n3
p9
b1
b2
b3
b4
b5
b6
b7
b8
b9
b10
b11
b12
b13
b14
b15
b16
(b)
n1
n2
p6
p2
p3
p1
p4
p5
n4
n3
b1
b2
b3
n5
(c) I3 - MN with C25%(c) I3 - MN with C25%
n1
n2
n3
n4
n5
n6
n8
n9
n7
p2
p1
p3
p4
b5
b4
b1
b2
b3
b6 b8
b7
b9
b10
b11
(d) I4 - MN with C50%(d) I4 - MN with C50%
n1
n2
p2
p3
p1
p4
n4
n3
b1
b2
n5n6
(e)
n1
n2
n3
n4
n5
n6
n8
n9
n7
p2
p1
p3
p4
b5
b4
b1
b2
b3
b6
n10
p5
p6
p8
p7
p9
p10
(f)
Figure 6: Artificial maps.
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 16 / 22

17. Computational Results
Table 1: Results obtained after evaluating different strategies in artificial maps.
Intel i5 Intel Edison
Methods SC
(ms)
Ψb Ψm Avg Time
(ms)
Ψb Ψm Avg Time
(ms)
GH - 86.7% 60.0% 73.3% 41 86.7% 60.0% 73.3% 347
GA 250 96.3% 72.0% 84.2% 250 72.3% 62.3% 67.3% 250
GA 500 99.0% 73.0% 86.0% 500 84.7% 65.3% 75.0% 500
GA 1000 98.3% 75.7% 87.0% 1000 91.0% 68.0% 79.5% 1000
GA-GH 250 95.3% 71.0% 83.2% 250 89.3% 62.7% 76.0% 309
GA-GH 500 100.0% 72.3% 86.2% 500 90.3% 65.7% 78.0% 510
GA-GH 1000 99.3% 76.0% 87.7% 1000 92.3% 69.0% 80.7% 1000
GA-GA 250 99.3% 72.3% 85.8% 250 81.0% 65.0% 73.0% 250
GA-GA 500 99.7% 73.3% 86.5% 500 89.7% 68.3% 79.0% 500
GA-GA 1000 99.7% 78.0% 88.8% 1000 94.7% 70.0% 82.3% 1000
Final Avg - 97.4% 72.4% 84.9% - 87.2% 65.6% 76.4% -
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 17 / 22

18. Computational Results
F3
F2
F4
F1
Legend:
route motor failure
route battery failure
critical situation detected
start emergencial route
critical distance
Figure 7: Results of case study in a real world scenario using a GA method.
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 18 / 22

19. Computational Results
Table 2: Results obtained after evaluating different strategies in the study case.
Intel i5 Intel Edison
Methods SC
(ms)
Ψb Ψm Avg Time
(ms)
Ψb Ψm Avg Time
(ms)
GH - 25.0% 0.0% 12.5% 22 25.0% 0.0% 12.5% 192
GA 250 100.0% 50.0% 75.0% 250 35.0% 37.5% 36.3% 250
GA 500 100.0% 50.0% 75.0% 500 62.5% 50.0% 56.3% 500
GA 1000 100.0% 67.5% 83.8% 1000 95.0% 50.0% 72.5% 1000
GA-GH 250 100.0% 50.0% 75.0% 250 37.5% 37.5% 37.5% 234
GA-GH 500 100.0% 50.0% 75.0% 500 67.5% 50.0% 58.8% 481
GA-GH 1000 100.0% 75.0% 87.5% 1000 92.5% 50.0% 71.3% 992
GA-GA 250 100.0% 50.0% 75.0% 250 50.0% 45.0% 47.5% 250
GA-GA 500 100.0% 50.0% 75.0% 500 85.0% 50.0% 67.5% 500
GA-GA 1000 100.0% 75.0% 87.5% 1000 100.0% 50.0% 75.0% 1000
Final Avg - 92.5% 51.8% 72.1% - 65.0% 42.0% 53.5% -
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 19 / 22

20. Computational Results
Figure 8: Complete route obtained in the SITL experiment.
https://youtu.be/k38GBjM5jXU
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 20 / 22

21. Conclusions
As expected, the personal computer leverages the strategies
performance.
GA-GA presents a slightly better performance.
This indicates the robustness of GAs to find good solutions despite
hardware limitations.
GA can be embedded and aid the decision making during fully
autonomous flights.
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 21 / 22

22. Acknowledgements
Email to contact:
{jesimar.arantes, andrebm }@usp.br
{marcio, simoes, claudio}@icmc.usp.br
Thank You!
J. S. Arantes et al. (USP) ICTAI 2017 November – 2017 22 / 22