Architectural study of a Self-driving car

Group 9
SYSEN 5400 / MAE 5950

Pau Garcia Buzzi (pg428)

Dalton Kujawa (jdk277)
Owen (Yiwen) Hua (yh686)
Imran Khan (iak26)

11/28/2016
 Contents
• Introduction
• Stakeholder analysis
• Concept exploration and selection
• Functional architecture
• Architecture Enumeration
• Cost, schedule and risk metrics
• Performance metric
• Tradespace analysis
• Architecture Optimization
• Conclusion/Results
 Introduction
• A self-driving or driver less car is a vehicle that is capable of sensing its
environment and navigating without human input.
 Introduction
Components installed on the current Google Car.
 Stakeholder analysis
• Main stakeholders are

• Customer (driver/Owner) • Employees

• Regulators/NGOs/Government • Marketing
• Suppliers • Competitors
• Investors/Shareholders • Insurance companies
• Oil/Electrical companies • Infrastructure investment
• People
Stakeholder analysis
 System Goals
Goal Goal statement
An autonomous car The car must drive itself without commands from the passengers
The autonomous vehicle must not cause crashes, property damage, or
Safety algorithm
endanger lives - either directly or indirectly
Google car should have comfortable and adjustable seats. Google car
Spatial and comfortable should also create more interior space for the passengers to use as
interior space
workstations or for other activities

Comfortable driving technique The vehicle should be programmed to provide a comfortable ride when
driving in various conditions (altitude, slope, wet pavement).
The car should efficiently spend fuel in order to save customer money and
Fuel efficiency
protect the Environment
Google car should obey current traffic rules, identify traffic lights and signs,
Following traffic rules and act like driven by a real driver. Car should also be able to avoid
pedestrians.
Low cost The car should be cheap enough to be bought by the target demographic.
No illegal use The function of the autonomous should be limited and audited.
Adapt the current road condition Google Car must be able to identify the external road condition.
Mass production Google car should be ready to be mass produced and Commercialized
 Concept Exploration
• Minimal, General functions that the product must be
capable of:
– Lifting
– Propelling
– Sensing
– Guiding
– Processing
• These functions are decomposed within the
Functional Architecture step and used to determine
decisions in the Enumeration step.
Functional Architecture
 Functional Architecture
• DMM Mapping of functions to goals
 Architecture Enumeration
Decision Name Canonical Type Options / Elements
Size of the car Standard Form Sports car / Sedan / Familiar / Bus
Combustion / Diesel / Electric /
Type of motor Standard Form Compressed Natural Gas / Hydrogen /
Hybrid
Sensor selection Down Selecting Laser / Ultrasonic / Camera / Radar
Commercial / Customized /
In-vehicle computer selection Down Selecting Microcontrollers & computers
system / Cloud computer
Choice of methods to actuate
steering, brake, throttle Standard Form Direct / Indirect control

Screen / Touch Screen / Voice / Short-

Communication Methods Down Selecting length / Middle-length / Long-length
Damage Minimization Down Selecting Passenger / Pedestrians / Vehicle
Private owner / Sell to government /
Marketing Strategy Down Selecting U.S. Market / Global Sale / Direct Sale

4*6*2*(2^4) *(2^4) *(2^6) *(2^3) *(2^5) = 201326592 architectures

 Enumeration Constraints
1.A sports car cannot have neither a Hydrogen nor a CNG engine.
2.A bus cannot have electric motor.
3.When more than 1 sensor is used in the vehicle, there should be
at least two computer or microcontrollers receiving signals from
the sensor to create some redundancy.
4.An online (cloud streaming) computer cannot control end effector
directly. A controller that can receive mechanical input must be
added to the control mechanism.
5.We cannot have architectures with no sensors.
6.We cannot have architectures with no computers.
7.We cannot have architectures with no communications systems.
 Reference Architectures, Deterministic and
Random Sampling
The 5 reference architectures below represent the various packages that would be
provided to be purchased by consumers. They discern the corners of the tradespace and
any architecture generated can be evaluated through comparison to these reference
architectures.

• Reference Architecture 1: Basic Package

• Reference Architecture 2: Fully Loaded Package

• Reference Architecture 3: Family Package

• Reference Architecture 4: Public Transport Package

• Reference Architecture 5: Luxury Package

 Reference Architectures, Deterministic and
Random Sampling (II)
- Size of the car: Standard form with 4 optionsà41
NOT FOUND IN
- Type of motor: Standard form with 6 options. à61 DATABASE
- Sensor Selection: Down-selecting with 5 elements. à25
- Computer Selection: Down-selecting with 4 elements. à24 224 41 61  231 61 81
- Method of actuation: Standard form with 2 options. à21
- Communication Methods: Down-selecting with 6 elements. à26 27 Architectures
- Damage Minimization: Down-selecting with 3 elements. à23
- Marketing: Down-selecting with 5 elements. à25

The rest of architectures (112 – 5 – 27 = 80) are crated using

random sampling to make sure that:
– SF decisions are roughly evenly distributed.
– Every DS decision option is selected half of the time.
 Cost, schedule and risk metrics
• Development cost. This metric includes all costs from the development
and testing phase of the project. It is architecturally distinguishing since it
affects several decisions: computers and sensors (both in number and
type/complexity), actuator method, vehicle size and type of engine.

Reference Architecture Total Cost ($) = Coding + Prototype

Basic (with hybrid engine) 38910
Basic (with combustion engine) 35450
Family 46412
Public Transport 62170
Luxury 58092
Fully Loaded 57803
 Cost, schedule and risk metrics (II)
• Expected Loss. This metric allows for us to evaluate both total cost of the
project and programmatic risk . This metric is architecturally distinguishing
in five decisions: vehicle size, engine type, sensors used, computers used,
and damage minimization software used.

Package Expected loss ($)

Basic 50,022
Fully loaded 51,476
Family 50,184
Public transport 51,891
Luxury 51,456
Cost, schedule and risk metrics (III)
• Reliability . This metric allows us to evaluate the operational risk of the
system by determining the chance that our self-driving system operates
properly by the projected end of life of the vehicle 

Package Reliability

Basic . 946599854179350

Fully loaded . 904627500365676

Family . 946599854179350

Public transport . 951218648397240

Luxury . 951108315856322
 Performance metric

A self-driving or driver less car is a vehicle that is capable of

sensing its environment and navigating without human input.

• Capability of sensing
• Autonomous Driving Level
 Capability of sensing
• Situational analysis using various different sensor
 Capability of sensing

Sensor Type Sensing Detection Purpose

Distance (m) Angle
(degrees)
LIDAR 100 360 Detect
environment Type of Sensor FOV ()
with
accuracy
LIDAR 10000
Radar sensor 200 30 Monitor
speed of Radar Sensor 3333
other cars
Camera Sensor 125
Camera 30 50 Monitor
Ultrasonic Sensor 50*
Sensor distance of
other car
near by
Ultrasonic 0 0 Detect
Sensor motion of
Google car
 Autonomous Driving Level
• Level 0: Human does all the work
• Level 1: Cruise control
• Level 2: Control speed, turn, and stay in the lane (eg.
Tesla)
• Level 3: Make decisions like pass slow vehicle and change
lane
• Level 4: Act as self-driving in certain safe place such as
city downtown or college campus
• Level 5: Handle all driving tasks and go anywhere

To climb the ladder, it requires update in sensors, computers (driving

algorithm), communication, driving actuation (no pedal, no steering wheel)
 Final Performance metric

By multiplying Autonomous driving level, the architecture which has

more required components (computer, actuation, communication
hardware) would be given a higher performance score.
 Tradespace Analysis
• Four metrics are used in the tradespace analysis - they are development
cost, expected loss, reliability, and combined performance
 Tradespace Analysis
• Sensitivity
Development Cost Expected Loss Reliability Sensing Capability Autonomous Driving Capability
LIDAR 8044.563 186.064 0.138 11710.015 0.273
Ultrasonic 1033.452 165.921 0.135 3388.003 0.019
Camera 1803.060 198.136 0.139 25.663 0.788
Radar 2452.366 168.814 0.143 105.293 0.784
Commercial 349.656 178.020 0.004 231.859 0.006
Custom 786.624 90.258 0.018 325.034 0.042
Microcontrollers 4420.521 366.952 0.048 285.367 0.069
Cloud 3439.514 266.434 0.024 68.524 0.013
Actuator 113.564 5.895 0.005 72.045 0.009
Screen 548.256 173.243 0.036 15.267 0.024
Physical 558.345 172.437 0.038 74.666 0.017
Voice 1375.638 354.785 0.032 4.697 0.032
Short 7512.731 7.032 0.040 201.286 0.005
Medium 3704.503 14.487 0.034 190.970 0.004
Long 1037.353 19.138 0.038 84.642 0.022
Passenger 533.081 384.155 0.002 76.129 0.001
Pedestrian 357.224 203.170 0.004 203.893 0.006
Vehicle 340.011 1082.742 0.001 131.465 0.006
Private 49.377 19.176 0.005 1.508 0.023
Government 96.835 36.134 0.001 137.946 0.006
US 95.070 11.592 0.004 74.749 0.011
Global 263.409 10.606 0.003 8.752 0.013
Dealer 414.679 8.777 0.003 127.941 0.001
Engine Type 8038.858 173.136 0.006 309.265 0.011
Vehicle Type 11377.720 947.568 0.005 1352.433 0.012
 Tradespace Analysis
• Connectivity
Development Cost Expected Loss Reliability Sensing Capability Autonomous Driving Capability
LIDAR 789.821 36.650 0.040 197.622 0.062
Ultrasonic 523.858 35.705 0.040 185.610 0.023
Camera 450.546 28.066 0.038 79.242 0.064
Radar 623.427 34.687 0.039 152.034 0.063
Commercial 520.151 49.866 0.006 271.722 0.014
Custom 695.200 50.903 0.024 432.765 0.052
Microcontrollers 992.169 54.383 0.026 538.929 0.065
Cloud 731.204 33.868 0.016 205.198 0.026
Actuator 283.028 14.446 0.006 86.206 0.012
Screen 405.527 32.154 0.019 32.508 0.029
Physical 512.209 29.136 0.018 114.891 0.017
Voice 414.524 37.195 0.018 40.036 0.027
Short 479.296 13.668 0.020 179.526 0.009
Medium 499.607 22.148 0.020 206.939 0.007
Long 522.509 23.780 0.021 146.270 0.024
Passenger 513.709 39.883 0.004 121.968 0.004
Pedestrian 445.755 34.382 0.007 147.498 0.010
Vehicle 451.111 32.177 0.003 192.202 0.009
Private 120.500 28.369 0.007 41.074 0.019
Government 160.897 28.267 0.002 141.599 0.012
US 178.419 15.305 0.005 135.460 0.013
Global 405.394 20.357 0.005 38.498 0.018
Dealer 501.073 14.449 0.005 155.110 0.007
Engine Type 517.689 23.383 0.005 137.665 0.011
Vehicle Type 305.870 20.006 0.005 219.985 0.011
 Tradespace Analysis
• Order of Decisions evaluated through
Simmons’ decision view chart:
Low Connectivity, High Connectivity,
High Sensitivity High Sensitivity
• Vehicle Type • Sensor Selection
• Engine Type • Computer Selection

Low Connectivity, Low High Connectivity,

Sensitivity Low Sensitivity
• Actuator Selection • Communication Methods
• Marketing Plan • Safety Algorithm
 Multi-objective Optimization using GA
• Chromosome: ‘bitString’ of 28 digits.
• Initial Population: Reference architectures +
deterministic sampling + Random enumeration  large
exploration at the beginning of the search, high
exploitation at the end.
• Population size: 4*length(Chromosome)=112.
• Number of generations: 200.
• The repair_bits() code ensures that the architectures
created by the creation, crossover and mutation
operators do not output unfeasible architectures.
• The fitness function we used takes into account all 5
metrics (Cost, Loss, Reliability, Autonomous Driving
Level and Sensing Ability)
 GA vs. Tradespace Pareto fronts
• The Pareto front obtained during the
tradespace analysis does not contain any of
the architectures of the Pareto front obtained
with the Genetic Algorithm:
– In our case, GA is expected to generate a better
approximation of the true Pareto Front.
– With partial random enumeration we can miss
“good” unexplored areas of the decision space.
 Conclusion
• RESULT: Pareto front of 40 architectures
Architecture Cost Loss Reliability Sensor Score Autodrive Level

1 72378,2 49882,8363 0,95110832 12458 2

2 46596,5 51909,3322 0,95111368 3458 3

3 29062 49908,1605 0,90258049 13333 1

4 59391,5 50483,3169 0,95110832 15833 1

5 85005 52134,7664 0,95100336 18333 1

1. Small car with hybrid engine, all sensors (4), 3 computers, direct control of actuators, with simple screen and
equipped with short, medium and long length communications systems and pedestrian damage minimization.
2. Sedan car with electric engine, 3 sensors , 4 computers, indirect control of actuators, all communications
systems and Passenger/Vehicle/Pedestrian damage minimization.
3. Sedan car with combustion engine, 2 sensors , 3 computers, direct control of actuators, with simple screen and
Passenger/Pedestrian damage minimization.
4. Family car with electric engine, 3 sensors, 4 computers, direct control of actuators, with touching screen,
medium and long length communications systems and pedestrian damage minimization
5. Bus with Compressed Natural Gas engine, 3 sensors, 3 computers, indirect control of actuators, with simple
screen, voice recognition system, short and long length communication systems and pedestrian damage
minimization. *Note: this last architecture would probably have to be upgraded to the maximum security level
assuring damage minimization of passengers as well.