Probabilistic Robotics and Models of Gaze Control. José Ignacio Núñez Varela

Transcription

1 Probabilistic Robotics and Models of Gaze Control José Ignacio Núñez Varela Seminario del Posgrado de Ingeniería Eléctrica 22 noviembre 2013

2 Outline: Introduction to probabilistic robotics Introduction to the problem of gaze control Candidate models of gaze control Experiments and conclusions Research projects

3 Part I: Probabilistic Robotics This part is mainly based on Chapters 1 and 2 of the book: Thurn et al. Probabilistic Robotics, MIT Press, 2005.

4 Basic Model: Robotic systems are situated in the real world, perceive information on their environment through sensors, and manipulate through physical forces.

5 Planning Sensing Acting Asimo Honda Picture credits:

6 Intelligent robotics Learning Reasoning Decision-making Planning Understanding Common sense PR2 Willow Garage

7 Robots have to be able to accomodate the enormous uncertainty that exists in the physical world. Imagen:

8 What factors contribute to the robot's uncertainty? Imagen:

9 Robot Environments: Inherently unpredictable Uncertainty is high for robots operating in the proximity of people

10 Robot Environments: Well structured environment << uncertainty Not structured environment >>> uncertainty

11 Robot Sensors: Sensors are limited in what they can perceive E.g., physical limitations affect range and resolution Sensors are subject to noise Sensors can break

12 Robot Actuators: Motors are, at some extent, unpredictable Control noise, wear-and-tear, mechanical failure

13 Robot's Internal Models (software): All internal models of the world are approximate Model errors have often being ignored

14 Algorithmic Approximations (software): Robots are real time systems, thus limiting the amount of computation being carried out Algorithms need to be approximated

15 Uncertainty Robots are forced to act even though it doesn't have sufficient information to make decisions with absolute certainty As robots are now moving into the open world, uncertainty becomes a major issue

16 Managing uncertainty is possibly the most important step towards robust real-world robot systems - Thurn, Burgard and Fox

17 Probabilistic Robotics Key idea: Represent uncertainty explicitly using the calculus of probability theory Instead of relying on a single best guess, probabilistic algorithms represent information by probability distributions over a whole space of guesses

18 Mobile Robot Localization: The problem of estimating a robot's coordinates relative to an external reference frame (the robot is given a map) A probability density function over the space of all locations represents the robot's belief This belief is updated using the robot's sensors

19 Mobile Robot Localization:

20 Major Paradigms Mid-1970s Mid-1990s Model-based Probabilistic robotics robotics Mid-1980s Behaviourbased robotics

21 Model-based vs. Probabilistic Robotics: Model-based robotics require accurate models of the robot, environment, etc. Probabilistic robotics have weaker requirements on this accuracy Behaviour-based vs. Probabilistic Robotics: Behaviour-based robotics require accurate sensors Probabilistic robotics have weaker requirements on this accuracy

22 Part II: Gaze Control Imagen:

23 Gaze Control Jason Babcock Biological perspective Machine perspective icub.org

24 Why study gaze control?

25

26

27

28

29 Foveal Vision cellfield.ca Michael Land

30

31

32 Eye Movements Saccades Aim: Shift the fovea to obtain high resolution samples Rapid jump-like movements (900 /sec) Ballistic (trajectory cannot change) Stereotyped (follow the same pattern) Voluntary and involuntary

33 Saccade Sequence

34 We perform hundreds or even thousands of saccades every day!

35 How does the brain decide where to fixate next?

36 Active Vision Ilya Repin

37 Yarbus

38 Task and context determine where to fixate next

39 Vision and Action Mary Hayhoe

40 Uncertainty Reduction

41 What mechanisms a rational decision maker could employ to select a gaze location optimally, or near optimally, given limited information and limited computation time during the performance of a task? Engineering science goal How humans select the next gaze location? Human behavioural goal

42 Gaze Control Processes

43 icub Humanoid Robot icub.org

44 Two problems where to look gaze allocation

45 Pick & Place Task

46

47

48

49

50 Models of Gaze Control Based on uncertainty reduction (Uncertainty) Based on rewards and uncertainty (Rew+Unc) Based on rewards, uncertainty and gain (Rew+Unc+Gain)

51 One-step look ahead gaze control What would happen if I look at entity ei?

52 Uncertainty Reduction How much uncertainty is reduced if I look at entity ei? X

53 Reward and Uncertainty How much value am I expected to get after looking at entity ei? X

54 Reward, Uncertainty & Gain Which motor system would get more benefit if gaze is allocated to it? X

55 Experiments We characterise how task performance varies in terms of three environmental parameters: Reach/grasp sensitivity Observation noise Field of view Also compared against Random and Round Robin gaze strategies

56 Reach/Grasp Sensitivity

57 Observation Noise

58 Field of View

59 Conclusions Gaze control models that incorporate rewards and uncertainty seem to perform better The Rew+Unc+Gain scheme has, in general, the best overall performance An active visual search process should be integrated into the Rew+Unc+Gain strategy The Rew+Unc+Gain and Rew+Unc gaze schemes were able to reproduce behavioural human data

60 Coordinación de Módulos de Control Guiados Visualmente en un Marco de Toma de Decisiones para Robots Humanoides Cesar A. Puente Montejano (responsable) Juan C. Cuevas Tello José I. Núñez Varela Omar Vital Ochoa Francisco E. Martínez Pérez Omar Rodríguez González Rogelio Castillo Morquecho Octavio Rentería Vidales Imagen:

61

62 Imagen:

63 Contact information: Jeremy L. Wyatt

64 Procesamiento de Señales Biomédicas Carlos Soubervielle Montalvo Omar Vital Ochoa Juan C. Cuevas Tello José I. Núñez Varela Imagen: IPN

65 Thank You!! Botodesigns / Chen Reichert jose.nunez@uaslp.mx Website: