for Path Planning using CNNs Eike Rehder, Jannik Quehl and Christoph - - PowerPoint PPT Presentation

▶
for path planning using cnns
SMART_READER_LITE
LIVE PREVIEW

for Path Planning using CNNs Eike Rehder, Jannik Quehl and Christoph - - PowerPoint PPT Presentation

Jun 30, 2023 240 likes •811 views

Driving Like a Human: Imitation Learning for Path Planning using CNNs Eike Rehder, Jannik Quehl and Christoph Stiller eike.rehder@daimler.com IROS 2017 | 24.09.2017 KIT The Research University in the Helmholtz Association www.kit.edu

slide-1
SLIDE 1

KIT – The Research University in the Helmholtz Association

IROS 2017 | 24.09.2017

www.kit.edu

Driving Like a Human: Imitation Learning for Path Planning using CNNs

Eike Rehder, Jannik Quehl and Christoph Stiller

eike.rehder@daimler.com

slide-2
SLIDE 2

2 Eike Rehder | IROS 2017 | 24.09.2017

Introduction: Path Planning

slide-3
SLIDE 3

3

Short Review: Dijkstra‘s Algorithm

Eike Rehder | IROS 2017 | 24.09.2017

slide-4
SLIDE 4

4

Short Review: Dijkstra‘s Algorithm

Eike Rehder | IROS 2017 | 24.09.2017

Start Goal

Find shortest path from start to goal:

slide-5
SLIDE 5

5

Short Review: Dijkstra‘s Algorithm

Start ? ? ? ? ? Goal

Find shortest path from start to goal:

Assign edge costs, node costs, Start = 0 4 2 1 4 5 4 2 2

Eike Rehder | IROS 2017 | 24.09.2017

slide-6
SLIDE 6

6

Short Review: Dijkstra‘s Algorithm

Start 4 2 ? ? ? Goal

Find shortest path from start to goal:

Assign edge costs, node costs, Start = 0 Propagate and sum costs 4 2 1 4 5 4 2 2

Eike Rehder | IROS 2017 | 24.09.2017

slide-7
SLIDE 7

7

Short Review: Dijkstra‘s Algorithm

Start 4 2 7 6 ? Goal 4 2 1 4 5 4 2 2

Find shortest path from start to goal:

Assign edge costs, node costs, Start = 0 Propagate and sum costs Expand cheapest node

Eike Rehder | IROS 2017 | 24.09.2017

slide-8
SLIDE 8

8

Short Review: Dijkstra‘s Algorithm

Start 4 2 7 6 ? Goal 4 2 1 4 5 4 2 2

Find shortest path from start to goal:

Assign edge costs, node costs, Start = 0 Propagate and sum costs Expand cheapest node

Eike Rehder | IROS 2017 | 24.09.2017

slide-9
SLIDE 9

9

Short Review: Dijkstra‘s Algorithm

Start 4 2 7 6 ? Goal 4 2 1 4 5 4 2 2

Find shortest path from start to goal:

Assign edge costs, node costs, Start = 0 Propagate and sum costs Expand cheapest node

Eike Rehder | IROS 2017 | 24.09.2017

slide-10
SLIDE 10

10

Short Review: Dijkstra‘s Algorithm

Start 4 2 7 5 ? Goal 4 2 1 4 5 4 2 2

Find shortest path from start to goal:

Assign edge costs, node costs, Start = 0 Propagate and sum costs Expand cheapest node Re-assign minimum cost

Eike Rehder | IROS 2017 | 24.09.2017

slide-11
SLIDE 11

11

Short Review: Dijkstra‘s Algorithm

Start 4 2 7 5 7 Goal 4 2 1 4 5 4 2 2

Find shortest path from start to goal:

Assign edge costs, node costs, Start = 0 Propagate and sum costs Expand cheapest node Re-assign minimum cost

Eike Rehder | IROS 2017 | 24.09.2017

slide-12
SLIDE 12

12

Short Review: Dijkstra‘s Algorithm

Start 4 2 7 5 7 Goal 4 2 1 4 5 4 2 2

Find shortest path from start to goal:

Assign edge costs, node costs, Start = 0 Propagate and sum costs Expand cheapest node Re-assign minimum cost Trace back shortest path

Eike Rehder | IROS 2017 | 24.09.2017

slide-13
SLIDE 13

13

Short Review: Dijkstra‘s Algorithm

Find shortest path from start to goal:

Start Goal

Eike Rehder | IROS 2017 | 24.09.2017

slide-14
SLIDE 14

14

Short Review: Dijkstra‘s Algorithm

Find shortest path from start to goal:

Graph Edges

Eike Rehder | IROS 2017 | 24.09.2017

slide-15
SLIDE 15

15

Short Review: Dijkstra‘s Algorithm

Find shortest path from start to goal:

Obstacle

Eike Rehder | IROS 2017 | 24.09.2017

slide-16
SLIDE 16

16

Short Review: Dijkstra‘s Algorithm

Find shortest path from start to goal:

Eike Rehder | IROS 2017 | 24.09.2017

slide-17
SLIDE 17

17

Shortest Path with a CNN

Eike Rehder | IROS 2017 | 24.09.2017

slide-18
SLIDE 18

18

Finding the Shortest Path with a CNN

Assign edge costs, node costs, Start = 0

Eike Rehder | IROS 2017 | 24.09.2017

slide-19
SLIDE 19

19

Finding the Shortest Path with a CNN

Assign edge costs, node costs, Start = 0 Propagate

Eike Rehder | IROS 2017 | 24.09.2017

slide-20
SLIDE 20

20

Finding the Shortest Path with a CNN

Assign edge costs, node costs, Start = 0 Propagate

=

Eike Rehder | IROS 2017 | 24.09.2017

slide-21
SLIDE 21

21

Finding the Shortest Path with a CNN

Assign edge costs, node costs, Start = 0 Propagate and sum costs

= + 1 =

Eike Rehder | IROS 2017 | 24.09.2017

slide-22
SLIDE 22

22

Finding the Shortest Path with a CNN

Assign edge costs, node costs, Start = 0 Propagate and sum costs

= +1 +1 +1 +1 +0

Eike Rehder | IROS 2017 | 24.09.2017

slide-23
SLIDE 23

23

Finding the Shortest Path with a CNN

Assign edge costs, node costs, Start = 0 Propagate and sum costs

=

Eike Rehder | IROS 2017 | 24.09.2017

slide-24
SLIDE 24

24

Finding the Shortest Path with a CNN

Assign edge costs, node costs, Start = 0 Propagate and sum costs Re-assign minimum cost

= min

Eike Rehder | IROS 2017 | 24.09.2017

slide-25
SLIDE 25

25

Finding the Shortest Path with a CNN

Assign edge costs, node costs, Start = 0 Propagate and sum costs Re-assign minimum cost

= min

Eike Rehder | IROS 2017 | 24.09.2017

slide-26
SLIDE 26

26

Finding the Shortest Path with a CNN

Cost Non-Zero Padding (!) Transition Filters Transition Cost

+

Cost per Action min pool Updated Cost Replace

Eike Rehder | IROS 2017 | 24.09.2017

“Reinforcement learning via recurrent convolutional neural networks”, arXiv:1701.02392

slide-27
SLIDE 27

27

Finding the Shortest Path with a CNN

Cost Non-Zero Padding (!) Transition Filters Transition Cost

+

Cost per Action min pool Updated Cost Replace Argmin of this layer is transition policy

Eike Rehder | IROS 2017 | 24.09.2017

slide-28
SLIDE 28

28 Eike Rehder | IROS 2017 | 24.09.2017

Evaluating the Shortest Path with a CNN + *

Current State Transition Policy Transition Selection Flipped Transition Filters Next State argmin Destination State

slide-29
SLIDE 29

29

Example: Simple Path Planning

Eike Rehder | IROS 2017 | 24.09.2017

slide-30
SLIDE 30

30

Example: Path Planning

Start Goal Obstacle

Find shortest path from start to goal:

Eike Rehder | IROS 2017 | 24.09.2017

slide-31
SLIDE 31

31

Nine possible transition filters Cost is the traversed distance

Example: Path Planning

+1 +0 +√2 +1 +1 +1 +√2 +√2 +√2

Eike Rehder | IROS 2017 | 24.09.2017

slide-32
SLIDE 32

32

Example: Path Planning

Cost Model

Additive layer High cost where obstacle is located

Eike Rehder | IROS 2017 | 24.09.2017

slide-33
SLIDE 33

33

Example: Path Planning

Cost Map State Transition Map

Eike Rehder | IROS 2017 | 24.09.2017

slide-34
SLIDE 34

34

Finding the Shortest Path with a CNN

If you use Dijkstra:

Graph traversal with known transitions is faster States can be updated selectively Visited nodes will never be touched again

Why would you do it then?

Eike Rehder | IROS 2017 | 24.09.2017

slide-35
SLIDE 35

35

Driving Like a Human: Imitation Learning

Eike Rehder | IROS 2017 | 24.09.2017

slide-36
SLIDE 36

36

Imitation Learning

Eike Rehder | IROS 2017 | 24.09.2017

Arial view: Google Maps

Intersection in Karlsruhe

slide-37
SLIDE 37

37

Imitation Learning

Recorded trajectories Teach a network to imitate human behavior

Eike Rehder | IROS 2017 | 24.09.2017

Intersection in Karlsruhe

Arial view: Google Maps

slide-38
SLIDE 38

38

Cost Non-Zero Padding (!) Transition Filters Transition Cost

+

Cost per Action min pool Updated Cost Replace

Imitation Learning

Eike Rehder | IROS 2017 | 24.09.2017

slide-39
SLIDE 39

39

Imitation Learning

Cost Non-Zero Padding (!) Transition Filters Transition Cost

+

Cost per Action min pool Updated Cost Replace Fill in the whole bunch of CNN techniques

Eike Rehder | IROS 2017 | 24.09.2017

slide-40
SLIDE 40

40

Example II: Imitation Learning +

In our case: FC-ResNet operating

  • n the arial view

Eike Rehder | IROS 2017 | 24.09.2017

slide-41
SLIDE 41

41

Imitation Learning

Path driven by human

Eike Rehder | IROS 2017 | 24.09.2017

slide-42
SLIDE 42

42

Imitation Learning

Path driven by human Cost map from arial image

Eike Rehder | IROS 2017 | 24.09.2017

slide-43
SLIDE 43

43

Imitation Learning

Path driven by human Cost map after planning

Eike Rehder | IROS 2017 | 24.09.2017

slide-44
SLIDE 44

44

Imitation Learning

Path planned by network Cost map after planning

Eike Rehder | IROS 2017 | 24.09.2017

slide-45
SLIDE 45

45

Imitation Learning

Path planned by network Path driven by human

Eike Rehder | IROS 2017 | 24.09.2017

Cost map after planning

slide-46
SLIDE 46

46

Imitation Learning

Path planned by network Path driven by human

Eike Rehder | IROS 2017 | 24.09.2017

slide-47
SLIDE 47

47

Outlook: Prediction and Cooperation

Eike Rehder | IROS 2017 | 24.09.2017

slide-48
SLIDE 48

48

Outlook: Pedestrian Prediction

Camera image Semantic map and top view Teach a network to predict human motion by planning Road Sidewalk Obstacles

Eike Rehder | IROS 2017 | 24.09.2017

slide-49
SLIDE 49

49

Outlook: Pedestrian Prediction

Crop of map centered around pedestrian Road Sidewalk Obstacles Pedestrian

Eike Rehder | IROS 2017 | 24.09.2017

“Pedestrian Prediction by Planning using Deep Neural Networks”, arXiv:1706.05904

slide-50
SLIDE 50

50

Outlook: Pedestrian Prediction

Predict destination for planning Road Sidewalk Obstacles Pedestrian

Eike Rehder | IROS 2017 | 24.09.2017

“Pedestrian Prediction by Planning using Deep Neural Networks”, arXiv:1706.05904

slide-51
SLIDE 51

51

Outlook: Pedestrian Prediction

Predicted with Net Road Sidewalk Obstacles Pedestrian

Eike Rehder | IROS 2017 | 24.09.2017

“Pedestrian Prediction by Planning using Deep Neural Networks”, arXiv:1706.05904

slide-52
SLIDE 52

52 Eike Rehder | IROS 2017 | 24.09.2017

Outlook: Cooperative Planning

Teach a network resolve conflicts

“Cooperative Motion Planning for Non-Holonomic Agents with Value Iteration Networks”, arXiv:1706.05904

slide-53
SLIDE 53

53

Summary

Eike Rehder | IROS 2017 | 24.09.2017

slide-54
SLIDE 54

54

Summary

Planning Net… … for imitation … for prediction … for cooperation

Eike Rehder | IROS 2017 | 24.09.2017

slide-55
SLIDE 55

55

The People

Florian Wirth Destination Prediction Jannik Quehl Trajectory Data Maximilian Naumann Cooperative Planning

Eike Rehder | IROS 2017 | 24.09.2017