Introduction to Vision & Robotics

Transcription

1 Introduction to Vision & Robotics by Bob Fisher Introduction to Robotics Introduction Some definitions Applications of robotics and vision The challenge: a demonstration Historical highlights 1

2 Vision and Robotics: some definitions Connecting the computer to the raw unwashed world (Russell & Norvig) create [from 2 d image] an accurate representation tti of the three dimensional world and its properties, then using this information we can perform any visual task (Aloimonos & Rosenfeld) Vision is the direct extraction of affordances from the optic array Gibson A robot is: A programmable multi function manipulator designed to move material, parts, or specialised devices through variable programmed motions for the performance of a variety of tasks (Robot Institute of America) Robotics is the intelligent connection of perception to action (Brady) Applications: dull, dirty or dangerous Visual inspection of parts Detecting crime on CCTV Welding on cars N.B. Overlap with automation 2

3 Applications: dull, dirty or dangerous Robot vacuum cleaners Cleaning nuclear plants Robot sewer inspection N.B. Overlaps with teleoperation Applications: dull, dirty or dangerous Visual aids for driving Demining Space exploration 3

4 Applications: also...? Entertainment industry Service industry Science A challenging problem We donʹt have much introspective insight into how we see or how we control action Building vision and robot systems involves a variety of interacting technology domains: Mechanical, electrical, digital, computational... This has proved to be a hard problem for AI Can beat the human grandmaster at chess Canʹt replace a house cleaner 4

5 Vision and robotics uses all areas of AI: Problem solving, planning, search, inference, knowledge representation, learning etc... But we canʹt just plug sensors and effectors onto an AI simulation and expect it to work Have constraints such as: Limited, noisy, raw information Continuous dynamic problem space Time, power, cost and hardware limitations Often solutions grounded in these constraints do not resemble conventional AI approaches Ancient Greek hydraulic and mechanical automata Hero of Alexandria AD 100 5

6 Renaissance optics: The algorithmic connection between the world and the image - Durer c th century clockwork animals Vaucanson s duck 6

7 Early 20 th century Electronic devices for remote control Tesla Methods for transducing images into electrical signals Robot used to describe artificial i humanoid slaves in Capek s play Rossum s Universal Robots s 1950s Development of electronic computer and control theory Used for artificial creatures e.g. Walter s tortoise and John Hopkins beast 7

8 1960s Industrial robot arms: Unimation Methods for image enhancement and pattern recognition 1970s Work on systems in restricted domains e.g. Shakey in blocks world Freddy assembly task 8

9 1980s Tackling more realistic problems: Natural scene analysis Face recognition Dynamic locomotion Significant impact in manufacturing Active vision Recent highlights: Leg Lab - MIT 1980 onward 1995 biped acrobatics 9

10 (Leg lab continued) 2000 complex biped Recent highlights: NavLab CMU 1987 onwards 1995 No hands across America drive from Pittsburgh to SanDiego 98.2% autonomous 10

11 Military Predator UAV Walking Reactive Insects Atilla & Ghengis MIT Brooks Lab c

12 Barrett Gripper Classical Control Paradigm: SPA SPA is serial ad hoc analytical assumptous SPA lacks speed and efficiency flexibility and adaptivity modularity and scalability error-tolerance and robustness 12

13 Quotations by R. Brooks fast, cheap, and out of control Planning is just a way of avoiding figuring out what to do next. The world is its own best model Complex behavior need not necessarily be the product of a complex control system Simplicity is a virtue Robots should be cheap All on board computation is important Systems should ldbe build incrementally Intelligence is in the eye of the observer No representation, no calibration, no complex computers Subsumption Architecture if not if not Evaluation of progress Scheduling of subtasks (sub-) Goal approach if not Path planning if not Self-localization localization & -calibration if not Obstacle avoidance sense if not Move when clear act 13

14 Evaluation of the Subsumption Architecture I wouldn t want one to be my chauffeur (C. Torpe) Modifications at low levels affect higher levels Often there the hierarchy is not strict Priorities rather than inhibition Representations, plans, and models do help Reproducibility is a virtue SPA is top down, SubsArc is bottom up neats vs. scruffies Sensing the world Keypoints: Why robots need sensing Factors that affect sensing capability Contact sensing Proximity and range sensing Sensing light 14

15 Why robots need sensing For a robot to act successfully in the real world it needs to be able to perceive the world, and itself in the world. Can consider sensing tasks in two broad classes: Finding out what is out there: e.g. is there a goal; is this a team-mate; is there danger? = Recognition Finding out where things are: e.g. where is the ball and how can I get to it; where is the cliff-edge and how can I avoid it? = Location But note that this need not be explicit knowledge Sensing capability depends on a number of factors: 1 What signals are available? Light Pressure & Sound Chemicals 15

16 N.B. Many more signals in world than humans usually sense: e.g. Electric fish generate electric field and detect distortion Sensing capability depends on a number of factors: 1 What signals are available? 2 What are the capabilities of the sensors? Distance Vision Hearing Smell Contact Taste Pressure Temperature Internal Balance Actuator position/ movement Pain/damage 16

17 Note this differs across animals: e.g. Bees see ultraviolet light Need to choose what to build in to robot options and costs Visible More like a target? Ultraviolet Sensors perform transduction Transduction: transformation of energy from one form to another (typically, (yp y, into electrical signals) 17

18 Sensors perform transduction Sensor characteristics mean there is rarely an isomorphic mapping between the environment and the internal signal, e.g: - Most transducers have a limited range - Most transducers have a limited resolution, accuracy, and repeatability - Most transducers have lags or sampling delays - Many transducers have a non-linear response - Biological transducers are often adaptive - Good sensors are usually expensive in cost, power, size Sensing capability depends on a number of factors: 1 What signals are available? 2 What are the capabilities of the sensors? 3 What processing is taking place? E.g. extracting useful information from a sound signal is difficult: 18

19 Sound sources cause air vibration Diaphragm (ear drum or microphone) has complex pattern of vibration in response to sound Usually analysed by separating frequencies and grouping through harmonic/temporal cues frequency time Sensing capability depends on a number of factors: 1 What signals are available? g 2 What are the capabilities of our sensors? 3 What processing is taking place? 4 What is the task? 19

20 Classical view Transduction Processing Internal model Actuators Task Decision on Action Plan of Action Motor commands Alternative view Task specific Task specific Task specific transduction ti processing action Task 2 specific transduction Task 3 specific transduction Task 2 specific processing Task 3 specific processing Task 2 specific action Task 3 specific action Principal function is location E.g. bump switch or pressure sensor: is the object contacting this part of the robot? Contact sensors 20

21 Principal function is location E.g. bump switch or pressure sensor: is the object contacting this part of the robot? Antennae: extend the range with flexible element Contact sensors Contact sensors Can also use for recognition e.g. Is it moving or are you? Human touch can distinguish shape, force, slip, surface texture Rat whiskers used to distinguish textures 21

22 Contact sensors Note these kinds of sensors can also be used to detect flow e.g. wind sensors Proximity and Range Sensors Again main function is position: distance to object at specific angle to robot Typically works by emitting signal and detecting ti reflection Short range = proximity sensor, e.g. IR 22

23 Proximity and Range Sensors Over longer distance = range sensors e.g. Sonar: emit sound and detect reflection 23

24 a. Sonar reflection time gives range b. Can only resolve objects of beam width c. Apparent range shorter than axial range d. Angle too large so wall invisible e. Invisible corner f. False reflection makes apparent range greater Using sonar to construct an occupancy grid Robot wants to know about free Robot wants to know about free space Map space as grid Each element has a value which is the probability it contains an obstacle Update probability estimates from sonar readings 24

25 25 Learning the map Assuming robot knows where it is in grid, sensory input provides noisy information about obstacles, e.g. for sonar R β α r II I III s Probability p(z O) of grid element z=(r,α) in region I if occupied (O) given measurement s Using Bayesian approach where p(o) will depend on previous measurements + 2 β α β λ R r R ) (~ ) ~ ( ) ( ) ( ) ( ) ( ) ( O p O z p O p O z p O p O z p z O p + = Sample occupancy grid Noisy fusion of multiple sonar observations

26 Proximity and Range Sensors More accurate information (same principle) from laser range finder Either planar or scanning 1,000,000 pixels per second Range of 30 metres Accuracy of few mms Sample Laser Scan 26

27 Light sensors Why is it so useful to detect light? Straight lines mean the rays reflected from objects can be used to form an image, giving you where. Very short wavelengths gives detailed structural information (including reflectance properties of surface, seen as colour) to determine what. Very fast, it is especially useful over large distances. But requires half our brain to do vision Conclusions Robots need sensing: location, objects, obstacles Commonly used sensors: laser range, sonar, contact, proprioceptic, GPS (outdoors), markers General scene scanning vs affordances 27