Outline

1
Outline

• ALVINN
• Autonomous Land Vehicle in a Neural Network

2
Introduction

• ALVINN started in 1987 as a machine learning
  project using supercomputers
• The main goal is to design a system that is
  capable of learning to drive
• On different kinds of road types and
• Using different kinds of sensors
• In contrast, existing methods for autonomous
  driving at that time were designed manually
• The programmer determines which features are
  important, then develop detectors, and algorithms
  to determine steering direction

3
ALVINN
4
(No Transcript)
5
Basic ALVINN Network Architecture

• Input layer consists of a single 30 x 32 retina
• Which can be from either a video camera or a
  scanning laser range finder
• Hidden layer consists of 4 four hidden units
• The output layer consists of 30 output units
• A linear representation of the current
  appropriate steering direction

6
ALVINN Network Architecture

• Why using 30 output units?

7
Network Training

• Initially, the network was trained using
  artificially generated road images and the
  corresponding correct steering directions
• It achieved some limited success
• However, it is difficult to generate training
  images that are realistic to real-world road
  scenes

8
On-the-fly Training

• In theory, it should be possible to teach the
  network to imitate a person as he/she drives
• Using the current sensor image as input and the
  persons current steering direction as the
  desired output
• This should reduce the required human effort to
  develop networks for new situations and allow the
  system to adapt quickly to new situations

9
On-the-fly Training
10
On-the-fly Training cont.

• Potential Problems
• The network will never be presented with
  situations where it must recover from
  misalignment errors
• The person only drives at the center of the road
  during training
• Overtraining of recent input
• For example, if the person drives the Navlab down
  a long segment of straight road at the end of the
  training, the network may forget what it had
  learned about driving on curved roads

11
On-the-fly Training cont.

• One solution that seems reasonable is to have the
  driver swerve the vehicle during training
• The idea is to teach the network how to recover
  from mistakes by showing examples
• Will this work? Why or why not?

12
On-the-fly Training cont.

• A solution adopted is to transform each sensor
  image by shifting and rotating to create
  additional images in which the vehicle may be

13
(No Transcript)
14
(No Transcript)
15
Extrapolating Missing Pixels
16
Extrapolating Missing Pixels
17
Transforming the Steering Direction
18
Buffering

• To increase diversity, previously encountered
  training patterns are buffered
• When new patterns are acquired, which old
  patterns should be replaced?
• When choosing the pattern to replace, a pattern
  whose replacement will bring the average steering
  direction closest to straight will be replaced

19
Effectiveness of Transformations and Buffering
20
Network Confidence

• IRRE Input reconstruction reliability
  estimation
• A measure of the familiarity of the input image
  to the network
• This is done by using an additional set of output
  units

21
(No Transcript)
22
On-the-fly Training cont.

• Different roads that ALVINN was able to handle
  after training

23
Network Representation Analysis
24
(No Transcript)
25
Driving with Alternative Sensors

• Night driving using laser reflectance images
• Contour following and obstacle avoidance using
  laser range images

26
Tactical Driving

• ALVINN was initially used as a stand-alone lane
  keeping system or a lane departure warning system
• As the system became mature, it is desirable to
  execute tactical driving including changing lanes
  and navigating intersections

27
(No Transcript)
28
Dual-View Land Transition Technique
29
(No Transcript)
30
(No Transcript)
31
(No Transcript)
32
(No Transcript)
33
(No Transcript)
34
(No Transcript)
35
(No Transcript)
36
Obstacle-Avoidance Maneuver
37
Intersection Navigation
38
(No Transcript)
39
(No Transcript)
40
(No Transcript)
41
(No Transcript)
42
(No Transcript)
43
(No Transcript)
44
Rule-based Multi-network Arbitration

• While individual networks are capable of
  performing aspects of autonomous driving, driving
  requires more than a collection of isolated
  capabilities
• To achieve better autonomy, a system must
  determine which capabilities should be employed
• This is achieved by training multiple networks
  and arbitrate among them through a rule-based
  module

45
Annotated Map System
46
Integrated ALVINN Architecture
47
Other Applications

• The techniques are very flexible
• The same techniques are used to guide a space
  robot to walk on the exterior of space station
  Freedom
• Named Self-Mobile Space Manipulator (SM2)

48
Limitations and RALPH

• While ALVINN is successful, the biggest
  limitation is relatively long training time
• RALPH (Rapidly Adapting Lateral Position Handler)
  was developed using three steps
• Sample the input
• Determine the road curvature
• Determine the lateral offset and steering

49
RALPH
50
RALPH cont.
51
Automated Highway System
52
1997 Free Agent Demonstration
53
Summary

• ALVINN is one of the most successful neural
  network applications
• By learning from inputs, it achieves flexibility
  that methods based on design do not have
• The same network can be used to learn driving
  under different road conditions using different
  sensors
• It has driven Navlab for several thousands of
  miles
• It set several world records