Autonomous driving changes our mobility. Autonomous driving has been around for years, is it really so? Is complete human driving behaviour to replace by autonomous driving and how? This talk address these challenges. It points out which technologies can be used. It address the intelligent and autonomous decisions in real time and show how a few of these capabilities can look like.
1. Davor Andric, Proprietary and Confidential
Autonomous driving: Reality or
fiction?
Davor Andric,
CTO, DXC Analytics for NCE
Founder, CTO, RobotsGoMental
25. Sep 2017
All opinions expressed here are my own.
2. Davor Andric, Proprietary and Confidential 2April 15, 2018
DAVOR ANDRIC, Founder, Chief Technical Officer, Developer
An accomplished computer science technologist, an architecture leader and a passionate
developer of AI technology and products. Over the last 20 years, Davor has been working
in the consulting, software and technology. His expertise is in designing and building
scalable platforms for analytics, machine learning and AI, and building products on those
platforms. He knows how to run a large technology project and bring it to a successful
end. He is great in working with both, people and machines.
All opinions expressed here are my own.
3. Davor Andric, Proprietary and Confidential 3April 15, 2018
Agenda
1. Challenges of autonomous driving
2. Current approaches to autonomous driving
3. Behavior of a human driver (Spencer vs Fast & Furious)
4. Is AI the big solution for autonomous driving?
(What is needed to build autonomous driving?)
5. How to act intelligently and autonomously in real time?
5. Davor Andric, Proprietary and Confidential 5April 15, 2018
Goals & Challenges
A big data & analytics platform is needed to effectively develop reliable algorithms for HAD.
360°
Object recognition
Situation analytics
Driving strategy
Potentially all data
are relevant
Pictures: Spiegel Online
Example:
Assumptions:
• 10 vehicles growth / yr
• 50TB / day
• 120 avg days of test /yr
• 2 dev-locations
0
100
200
300
400
500
600
700
2017 2018 2019 2020
PB
EB !
Full raw trace is
required
Data are captured
too fast to be
uploaded
Test drives are de-
centralized
6. Davor Andric, Proprietary and Confidential 6April 15, 2018
Some real hard challenges in autonomous driving
120 exabyte
Data
More than 10
complex
capabilities needed
▐ The legal road approval is far away or even impossible by using the
deterministic approach based on business rules
▐ Product development on this level is very hard or even impossible to
achieve with current approach
Three Examples:
22-25 mil.
km
1 mil.
Person
Years per
simple
capability
Amount of test data to train
the neuronal networks1
2 Effort to label the data
for feature recognition
3
7. Davor Andric, Proprietary and Confidential 7April 15, 2018
SAE* Automation Levels
*SAE International is a global association of more than 128,000 engineers and
related technical experts in the aerospace, automotive and commercial-vehicle industries.
9. Davor Andric, Proprietary and Confidential 9April 15, 2018
Current approach
• Platform
• ML & Deterministic Models on different
Software Development Kits (SDK)
• System
• Hardware
Current approaches to autonomous driving: Deterministic
Programming Models + Neural Networks
10. Davor Andric, Proprietary and Confidential 10April 15, 2018
Different ML models are used to build perceptual functions
similar to the visual system of a human driver.
Tree-Structured Models
Semantic Stixels Models
Faster SSD, RCNN, R-FCN
Source: Google Research
Source: Tree-Structured Models, Cordts, Rehfeld
Source: Semantic Stixels: Depth is Not Enough, Schneider, Cordts
11. Davor Andric, Proprietary and Confidential 11April 15, 2018
Example of motion control
• Structured Programming
Model*
• Deterministic Rules
• Neural Networks
* Source diagram dos not contain neural networks
12. Davor Andric, Proprietary and Confidential 12April 15, 2018
Current solution for motion control: Estimating steering accuracy
• In a very simplified view, the probability of finite or
countable event set to „achieve right steering“ in motion
control can be defined with probability theory:
• That is, the probability function f(x) lies between zero and
one for every value of x in the sample space Ω, and the
sum of f(x) over all values x in the sample space Ω is
equal to 1. An event is defined as any subset E of the
sample space Ω. The probability of the event E is defined
as
P(E) = 0,86 * 0,92 * 0,98 * 0,89 = 0,69 = 69%
This is based on fictive values and should serve as an illustration.
E1
E2
E3
E4
13. Davor Andric, Proprietary and Confidential 13April 15, 2018
The problems with current solutions based on machine
learning (ML) & deterministic rules:
1. Not really intelligent – Human intelligence doesn’t need 1000’s of images of a
digit three to be able to recognize it, unlike AI agents trained with current
machine learning.
2. Not really excellent - the ability to achieve the required accuracy based on the
current solution is not given.
3. Poor dealing with complexity – deterministic rules are usable in closed
environments such as a railway system. Such rules can't describe the
complexity of a real world.
4. Data volumes - for ML to work, enormous amounts of data are needed, which
are often not available or are too expensive.
5. Resources (Time & CPU) – A lot of time and power is needed to process big
volumes of data and learn from them, which is often not even feasible.
14. Davor Andric, Proprietary and Confidential 14April 15, 2018
ML with Rules is good to solve isolated use cases such:
following another car, assistance in breaking, parking.
ML with Rules are good in solving small isolated problem!
But they cannot describe the entire real word or
real events within which human drivers operate.
Current approach do not ensure full
autonomous driving.
The most important feature is missing!!!
The ability to rapidly adapt, learn and evolve.
That's AI.
AI
16. Davor Andric, Proprietary and Confidential 16April 15, 2018
FAIL FAST &
EVOLVE
Object
Detector
Object
Identifier
Object Tracker
Occupancy
Analyzer
Situation &
Motion
Analyzer
Drive
NavigatorMotion Actor
Loan & Street
Graph
Motion
Predictor
17. Davor Andric, Proprietary and Confidential 17April 15, 2018
AI is essential for autonomous driving
Driving like
a human
would
Act in real-time
High level of autonomy
Percept, Collect, Store,
Learn, Evolve, Act
TensorFlow, Coffee
AI platforms using
BigData & Analytics
Trained & abstracted
knowledge
Topology or
execution graphs on
Storm, Aka, Nifi,
TensorFlow
Intelligent driving agent
know how to drive
Required
New AI markets
are created
18. Davor Andric, Proprietary and Confidential 18April 15, 2018
How to act intelligently and autonomously in real time? …
based on theory on how life organizes, including the mind*
Inference through
standard ML models
& specialized agents
* Source: "Practopoiesis - a Theory on How Life Organizes, including the Mind" Danko-Nikolic (Aug 31, 2015)
T1
T2
T3
Applying fast learner
to small training
data sets
New AI technology /
science
Building learn-agents capable of learning the way
humans do
Training
takes
months
Fast learner from small training data sets
Knowledge
Training
takes weeks
Knowledge
Inference
100-200
ms
Training
takes secs