Lecture: Allows operation in enviroment without prior knowledge

Transcription

1 Lecture: SLAM

2 Lecture: Is it possible for an autonomous vehicle to start at an unknown environment and then to incrementally build a map of this enviroment while simulaneous using this map for vehicle localization - Tim Bailey Allows operation in enviroment without prior knowledge

3 Lecture: Is it possible for an autonomous vehicle to start at an unknown environment and then to incrementally build a map of this enviroment while simulaneous using this map for vehicle localization - Tim Bailey Allows operation in enviroment without prior knowledge Does not need access of independent position information, such as GPS

4 Lecture: Is it possible for an autonomous vehicle to start at an unknown environment and then to incrementally build a map of this enviroment while simulaneous using this map for vehicle localization - Tim Bailey Allows operation in enviroment without prior knowledge Does not need access of independent position information, such as GPS Example of SLAM problem formulations: Explore and return to starting point Traverse a region Find possible vehicle path to different goal locations Detect what changed in the environment

5 Lecture: SLAM Simultaneous What is SLAM acronyme of?

6 Lecture: SLAM Simultaneous Localization What is SLAM acronyme of?

7 Lecture: SLAM What is SLAM acronyme of? Simultaneous Localization and

8 Lecture: SLAM What is SLAM acronyme of? Simultaneous Localization and Map

9 Lecture: SLAM What is SLAM acronyme of? Simultaneous Localization and Map Building

10 Lecture: SLAM - Vocabular Meaning of some words: Localization - Determine the pose given a map

11 Lecture: SLAM - Vocabular Meaning of some words: Localization - Determine the pose given a map Mapping - Generate a map when pose is known

12 Lecture: SLAM - Vocabular Meaning of some words: Localization - Determine the pose given a map Mapping - Generate a map when pose is known SLAM - key steps Defined by an arbirary coordinate system (inital pose) Generate a map using sensors, and at the the same time compute pose Map errors and pose estimate are correlated

13 Lecture: SLAM - Vocabular Meaning of some words: Localization - Determine the pose given a map Mapping - Generate a map when pose is known SLAM - key steps Defined by an arbirary coordinate system (inital pose) Generate a map using sensors, and at the the same time compute pose Map errors and pose estimate are correlated Global Coordinate System

14 Lecture: SLAM - Vocabular Meaning of some words: Localization - Determine the pose given a map Mapping - Generate a map when pose is known SLAM - key steps Defined by an arbirary coordinate system (inital pose) Generate a map using sensors, and at the the same time compute pose Map errors and pose estimate are correlated Global Coordinate System Local Coordinte System (MAP)

15 Lecture: SLAM - Vocabular Meaning of some words: Localization - Determine the pose given a map Mapping - Generate a map when pose is known SLAM - key steps Defined by an arbirary coordinate system (inital pose) Generate a map using sensors, and at the the same time compute pose Map errors and pose estimate are correlated Global Coordinate System Local Coordinte System (MAP) Features - The components found in the mapo

16 Lecture: SLAM - Who defined SLAM? Professor Hugh F. Durrant-Whyte CAS - Centre of Excellence for Autonomous Systems

17 Lecture: SLAM - Who defined SLAM? Professor Hugh F. Durrant-Whyte CAS - Centre of Excellence for Autonomous Systems ACFR - Australian Centre for Field Robotics, Sydney

18 Lecture: SLAM - Who defined SLAM? Professor Hugh F. Durrant-Whyte CAS - Centre of Excellence for Autonomous Systems ACFR - Australian Centre for Field Robotics, Sydney PhD in System Engineering, Uni. of Pennsylvania, 1986

19 Lecture: SLAM - Who defined SLAM? Professor Hugh F. Durrant-Whyte CAS - Centre of Excellence for Autonomous Systems ACFR - Australian Centre for Field Robotics, Sydney PhD in System Engineering, Uni. of Pennsylvania, 1986 MSE in System Engineering, Uni. of Pennsylvania, 1985

20 Lecture: SLAM - Who defined SLAM? Professor Hugh F. Durrant-Whyte CAS - Centre of Excellence for Autonomous Systems ACFR - Australian Centre for Field Robotics, Sydney PhD in System Engineering, Uni. of Pennsylvania, 1986 MSE in System Engineering, Uni. of Pennsylvania, 1985 BSc in Nuclear Engineering, Univ. of London, 1983

21 Lecture: SLAM - Who defined SLAM? Professor John J. Leonard Works with Navigation, Guidance, and Control

22 Lecture: SLAM - Who defined SLAM? Professor John J. Leonard Works with Navigation, Guidance, and Control Applications with underwater robotics, survey robots

23 Lecture: SLAM - Who defined SLAM? Professor John J. Leonard Works with Navigation, Guidance, and Control Applications with underwater robotics, survey robots Professor of Mechanical and Ocean Engineering

24 Lecture: SLAM - Who defined SLAM? Professor John J. Leonard Works with Navigation, Guidance, and Control Applications with underwater robotics, survey robots Professor of Mechanical and Ocean Engineering Massachusetts Institute of Technology, Cambridge

25 Lecture: SLAM - Who defined SLAM? Professor John J. Leonard Works with Navigation, Guidance, and Control Applications with underwater robotics, survey robots Professor of Mechanical and Ocean Engineering Massachusetts Institute of Technology, Cambridge PhD, Univeristy of Oxford, 1984

26 Lecture: SLAM - Who defined SLAM? Professor John J. Leonard Works with Navigation, Guidance, and Control Applications with underwater robotics, survey robots Professor of Mechanical and Ocean Engineering Massachusetts Institute of Technology, Cambridge PhD, Univeristy of Oxford, 1984 BSEE, University of Pennsylvania, 1987

27 Lecture: SLAM - MAPS Metric Maps Example of a Topological Map

28 Lecture: SLAM - MAPS Metric Maps Topological Maps Example of a Topological Map

29 Lecture: SLAM - MAPS Example of a Topological Map Metric Maps Topological Maps Or Topologial and Metric Maps

30 Lecture: SLAM

31 Lecture: SLAM - Algorithms Kalman Filters Different implementation of SLAM

32 Lecture: SLAM - Algorithms Different implementation of SLAM Kalman Filters Particle Filers ( Monte Carlo simulations)

33 Lecture: SLAM - Algorithms Different implementation of SLAM Kalman Filters Particle Filers ( Monte Carlo simulations) EKF SLAM

34 Lecture: SLAM - Algorithms Different implementation of SLAM Kalman Filters Particle Filers ( Monte Carlo simulations) EKF SLAM Fast SLAM 2.0

35 Lecture: SLAM - Algorithms Different implementation of SLAM Kalman Filters Particle Filers ( Monte Carlo simulations) EKF SLAM Fast SLAM 2.0 Graph SLAM

36 Lecture: SLAM - Algorithms Different implementation of SLAM Kalman Filters Particle Filers ( Monte Carlo simulations) EKF SLAM Fast SLAM 2.0 Graph SLAM Occupancy Grid SLAM

37 Lecture: SLAM - Algorithms Different implementation of SLAM Kalman Filters Particle Filers ( Monte Carlo simulations) EKF SLAM Fast SLAM 2.0 Graph SLAM Occupancy Grid SLAM 3D SLAM

38 Lecture: SLAM - Algorithms Different implementation of SLAM Kalman Filters Particle Filers ( Monte Carlo simulations) EKF SLAM Fast SLAM 2.0 Graph SLAM Occupancy Grid SLAM 3D SLAM 6DOF SLAM

39 Lecture: SLAM - Grid Based

40 Lecture: SLAM - Grid Based

41 Lecture: SLAM - KF Steps Feature based SLAM, input and outputs Given: Input, control signal U 1:k = {u 1,u 2,...,u k } Observations Z 1:k = {z 1,z 2,...,z k }

42 Lecture: SLAM - KF Steps Feature based SLAM, input and outputs Given: Input, control signal U 1:k = {u 1,u 2,...,u k } Observations Z 1:k = {z 1,z 2,...,z k } Create: Map of features M = {m 1 n,m 2,m 3,...,m k } Robot path (pose)

43 Lecture: SLAM - KF Steps Feature based SLAM, input and outputs Given: Input, control signal U 1:k = {u 1,u 2,...,u k } Observations Z 1:k = {z 1,z 2,...,z k } Create: Map of features M = {m 1 n,m 2,m 3,...,m k } Robot path (pose)

44 Lecture: Feature Based SLAM

45 13-1

46 Lecture: SLAM - Feature based SLAM?

47 Lecture: SLAM - Correlation

48 Lecture: Joint state with momentary pose State vector wih features and momentary pose ˆX k = One pose in state vector x posek feature 1 feature 2. feature N

49 Lecture: Joint state with momentary pose State vector wih features and momentary pose ˆX k = One pose in state vector Number of features N x posek feature 1 feature 2. feature N

50 Lecture: Joint state with momentary pose State vector wih features and momentary pose ˆX k = One pose in state vector Number of features N x posek feature 1 feature 2. feature N Number of features vary over time

51 Lecture: SLAM State vector with features and pose history ˆX k = Pose history in state vector x posek. x pose1 x pose0 feature 1 feature 2. feature N

52 Lecture: SLAM State vector with features and pose history ˆX k = Pose history in state vector Number of features N x posek. x pose1 x pose0 feature 1 feature 2. feature N

53 Lecture: SLAM - Steps KF Example of steps in Kalman filter SLAM

54 Lecture: SLAM - KF problems in implementations Problems with Kalman filtered based SLAM implementations Difficult to manage data association

55 Lecture: SLAM - KF problems in implementations Problems with Kalman filtered based SLAM implementations Difficult to manage data association Especially difficult if: the environment is cluttered the environment is dynamic the environment has structural similarities ( wrong association)

56 Lecture: SLAM - KF problems in implementations Problems with Kalman filtered based SLAM implementations Difficult to manage data association Especially difficult if: the environment is cluttered the environment is dynamic the environment has structural similarities ( wrong association) Wrong acceptance of observation can destroy the state vector (map)

57 Lecture: SLAM - KF problems in implementations Problems with Kalman filtered based SLAM implementations Difficult to manage data association Especially difficult if: the environment is cluttered the environment is dynamic the environment has structural similarities ( wrong association) Wrong acceptance of observation can destroy the state vector (map) Reject ambigous measurements! Don t rely on nearest neighbour in data association

58 Lecture: SLAM - KF problems in implementations Problems with Kalman filtered based SLAM implementations Difficult to manage data association Especially difficult if: the environment is cluttered the environment is dynamic the environment has structural similarities ( wrong association) Wrong acceptance of observation can destroy the state vector (map) Reject ambigous measurements! Don t rely on nearest neighbour in data association Problems can also occur if feature estimation becomes too strong

59 Lecture: SLAM - KF problems in implementations Problems with Kalman filtered based SLAM implementations Robot path and map are unknowns

60 Lecture: SLAM - KF problems in implementations Problems with Kalman filtered based SLAM implementations Robot path and map are unknowns Errors in map and pose estimates are correlated

61 Lecture: SLAM - KF problems in implementations Problems with Kalman filtered based SLAM implementations Robot path and map are unknowns Errors in map and pose estimates are correlated The observations needs to be mapped to known landmarks

62 Lecture: SLAM - KF problems in implementations Problems with Kalman filtered based SLAM implementations Robot path and map are unknowns Errors in map and pose estimates are correlated The observations needs to be mapped to known landmarks If new unique landmars and them to map

63 Lecture: SLAM - KF problems in implementations Problems with Kalman filtered based SLAM implementations Robot path and map are unknowns Errors in map and pose estimates are correlated The observations needs to be mapped to known landmarks If new unique landmars and them to map Wrong data associations can lead to divergence

64 Lecture: SLAM - Darpa Darpa Urban Challenge, 96km Self-driving cars (Darpa, 2005, 2007)

65 Lecture: SLAM - Darpa Darpa Urban Challenge, 96km Self-driving cars (Darpa, 2005, 2007) Unmanned Aerial Vehicles (UAV)

66 Lecture: SLAM - Darpa Darpa Urban Challenge, 96km Self-driving cars (Darpa, 2005, 2007) Unmanned Aerial Vehicles (UAV) Autonomous Underwater Vehicles (AUV)

67 Lecture: SLAM - Darpa Darpa Urban Challenge, 96km Self-driving cars (Darpa, 2005, 2007) Unmanned Aerial Vehicles (UAV) Autonomous Underwater Vehicles (AUV) Planetary Rovers (Mars, 1997)

68 Lecture: SLAM - Darpa Darpa Urban Challenge, 96km Self-driving cars (Darpa, 2005, 2007) Unmanned Aerial Vehicles (UAV) Autonomous Underwater Vehicles (AUV) Planetary Rovers (Mars, 1997) Domestic Robots

69 Lecture: SLAM - Darpa Darpa Urban Challenge, 96km Self-driving cars (Darpa, 2005, 2007) Unmanned Aerial Vehicles (UAV) Autonomous Underwater Vehicles (AUV) Planetary Rovers (Mars, 1997) Domestic Robots Field Robotics ( Mining, Forestry)

70 Lecture: SLAM - Example of AUV SeaBED robot under Antarctic Ice, WHO Autonomous Underwater Vechicle (AUV)

71 Lecture: SLAM - Example of AUV SeaBED robot under Antarctic Ice, WHO Autonomous Underwater Vechicle (AUV) Accurate 3D maps of Antarctic Sea Ice

72 Lecture: SLAM - Example of AUV SeaBED robot under Antarctic Ice, WHO Autonomous Underwater Vechicle (AUV) Accurate 3D maps of Antarctic Sea Ice Fitted with an upward-looking sonar

73 Lecture: SLAM - Example of AUV SeaBED robot under Antarctic Ice, WHO Autonomous Underwater Vechicle (AUV) Accurate 3D maps of Antarctic Sea Ice Fitted with an upward-looking sonar Map and measure underside of ice cover

74 Lecture: SLAM - Example of AUV SeaBED robot under Antarctic Ice, WHO Autonomous Underwater Vechicle (AUV) Accurate 3D maps of Antarctic Sea Ice Fitted with an upward-looking sonar Map and measure underside of ice cover Was driven in lawnmover pattern

75 Lecture: SLAM - Example with self driving cars Self driving car - Driver can relax in tedious situations Lot of SLAM in this (localization)

76 Lecture: SLAM - Example with self driving cars Self driving car - Driver can relax in tedious situations Lot of SLAM in this (localization) Upcoming technology to support driver in: highway driving automated parking collision avoidance (oncoming cars)

77 Lecture: SLAM - Example with self driving cars Self driving car - Driver can relax in tedious situations Lot of SLAM in this (localization) Upcoming technology to support driver in: highway driving automated parking collision avoidance (oncoming cars)

78 Lecture: SLAM - Example with self driving cars Self driving car - Driver can relax in tedious situations Lot of SLAM in this (localization) Upcoming technology to support driver in: highway driving automated parking collision avoidance (oncoming cars)

79 Lecture: SLAM - Example with self driving cars Self driving car - Driver can relax in tedious situations Lot of SLAM in this (localization) Upcoming technology to support driver in: highway driving automated parking collision avoidance (oncoming cars)

80 Lecture: SLAM - AGV Automated Guided Vehicles (AGV) Automation of warehouses

81 Lecture: SLAM - AGV Automated Guided Vehicles (AGV) Automation of warehouses Fork lifts and transportation vehicles

82 Lecture: SLAM - AGV Automated Guided Vehicles (AGV) Automation of warehouses Fork lifts and transportation vehicles Increases efficiency and reduces cost

83 Lecture: SLAM - AGV Automated Guided Vehicles (AGV) Automation of warehouses Fork lifts and transportation vehicles Increases efficiency and reduces cost Markers on floor, lasers, and visual markers

84 Lecture: SLAM - AGV Automated Guided Vehicles (AGV) Automation of warehouses Fork lifts and transportation vehicles Increases efficiency and reduces cost Markers on floor, lasers, and visual markers Uses fiber optic gyros for very accurate azimuth control

85 Lecture: SLAM - Questions? End of presentation: Questions?