Emotional Robots

1
Emotional Robots

• Paranoid androids or intelligent agents?
• Dr Will Browne - Cybernetics

2
Emotions
The question is not whether intelligent machines
can have any emotions, but whether machines can
be intelligent without emotions. Marvin Minsky
This talk aims to show whether emotions can be
useful to robots in creating intelligent agents
or whether emotions will be only superficial
for human-robot interaction a robot that
smiles when it opens the door for you
3
Plan

• Background science fiction, robots and AI
• Why use a cognitive model in a robot?
• Alternative intelligent approaches for robot
  control
• Emotions in animals and robots
• Experiments with emotional robots
• Discussion, conclusions and the future

• Background science fiction, robots and AI
• Why use a cognitive model in a robot?
• Alternative intelligent approaches for robot
  control
• Emotions in animals and robots
• Experiments with emotional robots
• Discussion, conclusions and the future

4
(No Transcript)
5
Mobile Robotics

• Dr Susan Calvin obtained her bachelor's degree at
  Columbia in 2003 and began graduate work in
  cybernetics.
• Asimov (1940)

6
Picture of Lt Commander Data
7
The Most Advanced Robot?
8
This 1100 spin Bosch machine is incredibly quiet
and positively high-end. It has everything you
would expect to find on a Bosch including
exclusive features like the 3D AquaSpa wash
system with Fuzzy Control.
9
Plan

• Background science fiction, robots and AI
• Why use a cognitive model in a robot?
• Alternative intelligent approaches for robot
  control
• Emotions in animals and robots
• Experiments with emotional robots
• Discussion, conclusions and the future

• Background science fiction, robots and AI
• Why use a cognitive model in a robot?
• Alternative intelligent approaches for robot
  control
• Emotions in animals and robots
• Experiments with emotional robots
• Discussion, conclusions and the future

10
Why it is Good to be a Cognitive System!
11
Its not just ducks!
12
Owen Holland - University of Essex
How could the agent achieve its task (or mission)?

• by being preprogrammed for every possible
  contingency? No
• by having learned the consequences for the
  achievement of the mission of every possible
  action in every contingency? No
• by having learned enough to be able to predict
  the consequences of tried and untried actions, by
  being able to evaluate those consequences for
  their likely contribution to the mission, and by
  selecting a relatively good course of action?
  Maybe

13
Latent Learning

• Latent learning has three stages
• Robot (or Rat) enters the maze and explores it
  without reward.
• Robot (or Rat) is then placed in one of the end
  zones (E,F) and given a reward
• Robot (or Rat) is then placed at start (S) of
  maze and must navigate in the shortest path back
  to the reward state.

14
Learning Classifier Systems
CONDITIONS
MATCH
ENCODING
INITIAL RULE BASE
PLAUSIBLY BETTER RULES GENERATED
SELECT
RULE DISCOVERY
TRAINING RULE BASE
EFFECT
CREDIT
FINAL RULE BASE
DECODING
ENVIRONMENT
ACTIONS
LEARNING CLASSIFIER SYSTEM
UCL 2006
15
Latent Learning

• Robots not very good at consistently turning at
  90
• Latent learning environment simplified to
  N,E,W,S, compass points.
• After a five-minute run a robot will start
  getting very close to one wall and will
  eventually get stuck against it!

16
Plan

• Background science fiction, robots and AI
• Why use a cognitive model in a robot?
• Alternative intelligent approaches for robot
  control
• Emotions in animals and robots
• Experiments with emotional robots
• Discussion, conclusions and the future

• Background science fiction, robots and AI
• Why use a cognitive model in a robot?
• Alternative intelligent approaches for robot
  control
• Emotions in animals and robots
• Experiments with emotional robots
• Discussion, conclusions and the future

17
Robot Embodiment
18
Troy Kelley US Army Research Laboratory
Symbolic
Attention is the highest level goal
Production system operates on memories
Results go to memory
Semantic network
Parallel processing all of the inputs simultaneous
ly
Subsymbolic
Subsymbolic processing
Camera inputs Laser inputs Sound inputs
SS-RICS
Stimuli
19
Cartesian Theatre

• A Homunculus1 watches the individual needs,
• Thus, Domination through global selection
• Dennett and Kinsbourne
• a centred locus in the brain Cartesian
  materialism, because it is the view one arrives
  at when one discards Descartes dualism but fails
  to discard the associated imagery of a central
  (but material) theatre where it all comes
  together.
• http//wwwcms.brookes.ac.uk/p0054139/Paper/Decide
  r8.htm
• 1.little man
• http//faculty.washington.edu/chudler/flash/hom.ht
  ml

20
Global Workspace Architecture
Murray Shanahan Imperial College London

• Multiple parallel specialist processes compete
  and co-operate for access to a global workspace
• If granted access to the global workspace, the
  information a process has to offer is broadcast
  back to the entire set of specialists

21
GWT and the Frame Problem
Murray Shanahan Imperial College London

• Both Fodor and Dennett seem to have a strictly
  serial architecture in mind when they
  characterise the frame problem

• This certainly looks computationally infeasible

22
Subsumption Architecture

• Subsumption architectures are hybrid

23
Example Architectures

• Distributed Architecture for Mobile Navigation
  (DAMN)
• (Rosenblatt 1995)
• Agents vote for the actions, and the action which
  receives the most votes is executed.
• Similar to Maximise Collective Happiness
• Allows voting for actions other than their first
  choice.
• Mobile Autonomous Robot Software (Mars)
• (Carnegie Mellon University, Veloso)

24
SOAR (1)

• Developed by Allen Newell and others
• Purely Symbolic General Cognitive Architecture
  Symbolic
• All information held in production rules
• Rules relating to the problem brought to working
  memory a decision process then decides which
  rule to use
• Learning achieved through the resolution of
  impasses

25
SOAR (2)
Taken from J. Lehman, J. Laird, P. Rosenbloom,
A Gentle introduction to SOAR 2006 update,
http//ai.eecs.umich.edu/soar/sitemaker/docs/misc/
GentleIntroduction-2006.pdf, 2006
26
ACT-R (1)

• Developed by John Anderson et al. at Carnegie
  Mellon University
• Hybrid Symbolic/Subsymbolic General Cognitive
  Architecture
• Procedural (production system) and Declarative
  (subsymbolic) Memory systems
• Goal Stack a stack of goals of the system
  only the top (current) goal visible
• Achieved goals form a new item in declarative
  memory
• Limited Attentional Resource

27
ACT-R (2)
Taken from M. C. Lovett, L. M. Reder, and C.
Lebiere, "Modeling Working Memory in a unified
Architecture," in Models of Working Memory
Mechanisms of active maintenance and executive
control, A. Miyake and P. Shah, Eds. Cambridge
Cambridge University Press, 1999, pp. 135-182.
28
Perspective Taking A tale of two systems
Alan Schultz Naval Research Laboratory

• ACT-R/S (Schunn Harrison, 2001)
• Our perspective-taking system using ACT-R/S is
  described in Hiatt et al. 2003
• Three Integrated VisuoSpatial buffers
• Focal Object ID non-metric geon parts
• Manipulative grasping/tracking metric geons
• Configural navigation bounding boxes
• Polyscheme (Cassimatis)
• Computational Cognitive Architecture where
• Mental Simulation is the primitive
• Many AI methods are integrated
• Our perspective-taking using Polyscheme is
  described in Trafton et al., 2005

29
David Noelle University of California
Computational Cognitive Neuroscience Models

• Healthy performance on frontal tasks.
• Prolonged frontal developmental period.
• Monkey lesion data.
• Human frontal damage patient performance.
• Autistic peformance.

(Rougier, Noelle, Braver, Cohen, and O'Reilly,
2005)
30
David Noelle University of California
Robotic Working Memory
Could robot control systems benefit from the
inclusion of a working memory system?

• The highly limited capacity of working memory,
  along with its tight coupling with deliberation
  mechanisms, might alleviate the need for costly
  memory searches.
• Information needed to fluently perform the
  current task is temporarily kept handy in the
  working memory store.

Can computational neuroscience models of the
working memory mechanisms of the human brain shed
light on the design of a robotic working memory
system? The adaptive working memory toolkit
(WMtk) in C
31
Owen Holland University of Essex
Were trying to build a robot that has an
internal model of itself and an internal model of
the world, and that uses them to predict the
outcomes of novel or untried actions. And maybe
the IAM will be conscious
32
Plan

• Background science fiction, robots and AI
• Why use a cognitive model in a robot?
• Alternative intelligent approaches for robot
  control
• Emotions in animals and robots
• Experiments with emotional robots
• Discussion, conclusions and the future

33
Emotions - Neuroscience
Rolls, defines emotions as the states elicited
by reward and punishers, including changes in
those rewards and punishments, by which a
reward is anything for which an animal will work,
i.e. a Positive Reinforcer and a punishment is
anything that an animal will work to escape or
avoid, i.e. a Negative Reinforcer. Damasio,
stimulus good/bad evaulation Emotions are based
on experiences and help guide future actions.
34
(No Transcript)
35
Extending attention model with amygdala
John Taylor Kings College London

• Simulate by extending attention model with
  amygdala

36
ISAC by Kaz Kawamura
Short-term Memory Working Memory System Long-term
Memory
Haikonens System Reactions Theory of
Emotions Sensations - Reactions
37
http//info.fysik.dtu.dk/Brainscience/people/rodne
y.html
38
Other Emotional Architectures

• ASD (Maes 1990) Action selection dynamics
• ALEC
• DARE (Marcia et al. 2001)
• EBII
• AD with ECS (Malfaz et al.) emotional control
• Motives (goals) interact with beliefs
  (predictions) to produce emotions. Cf. Sloman
• Feelings arise after enaction of emotions.

39
Plan

• Background science fiction, robots and AI
• Why use a cognitive model in a robot?
• Alternative intelligent approaches for robot
  control
• Emotions in animals and robots
• Experiments with emotional robots
• Discussion, conclusions and the future

• Background science fiction, robots and AI
• Why use a cognitive model in a robot?
• Alternative intelligent approaches for robot
  control
• Emotions in animals and robots
• Experiments with emotional robots
• Discussion, conclusions and the future

40
Augmented Reality
41
Augmented Reality Pacbot demo
42
Augmented Reality Emotional Architecture
43
Augmented Reality Emotional Architecture
44
Augmented Reality Normal Explore
45
Augmented Reality Tight Explore
46
Augmented Reality Emotional Explore
47
Augmented Reality Explore Task
48
Explore efficiency
49
Explore effectiveness
50
Augmented Reality Tight Explore
51
Augmented Reality Very Tight Explore
52
Plan

• Background science fiction, robots and AI
• Why use a cognitive model in a robot?
• Alternative intelligent approaches for robot
  control
• Emotions in animals and robots
• Experiments with emotional robots
• Discussion, conclusions and the future

• Background science fiction, robots and AI
• Why use a cognitive model in a robot?
• Alternative intelligent approaches for robot
  control
• Emotions in animals and robots
• Experiments with emotional robots
• Discussion, conclusions and the future

53
Discussions
Visible Emotions shown useful in human-robot
interaction and should speed up robot-robot
interaction - However these can be top-down
interpretations of internal deterministic
states Internal emotions emerge from embodiment
and interaction with an environment - Reactive
mechanisms make possible alarm-driven primary
emotions while Deliberative mechanisms make
possible secondary emotions using global alarm
mechanisms linked to deliberative processes.
Aaron Sloman
54
Discussions

• What makes emotions useful?
• Emerge rather than hard coded
• Generalise across known and unknown situations
• Episodic and temporal
• Fast response if necessary
• Non-linear, non-deterministic and stochastic
• Emotions act as a warehouse where items can be
  sorted, grouped and sent appropriately.
• Rather than every individual manufacturers
  (internal external states) delivering to each
  outlet shop (effective action)

55
Discussions

• What makes emotions useful?
• Practically
• Much less lines of code for better behaviour -
  40
• Easier to understand code for non-programmers
• Intuitive behaviours result
• Difficulties
• Emotions levels require tuning
• Diagnosis and prediction of behaviour difficult
• When is it justified to call a non-linear
  controller an emotion?

56
Concluding Remarks

• Robots need real emotions to successfully
  complete complex real-world tasks.
• Emotions can set goals balance explore vs
  exploit
• Emotions can modify existing behaviours
• Emotions facilitate action in unknown domains

57
Future Cognitive Robots?

• Dont make cognition hard for ourselves
• Models are useful,
• but the mind is not so clear-cut
• Human cognition is a good model,
• but desired behaviour may be achieved by other
  models
• Increasingly powerful tools assist in advancing
  cognitive robotics, e.g., computational power,
  engineering materials and neurological
  understanding
• Emotions will play an important part in both
  Human-Robot interaction and enabling Autonomous
  Robot behaviour

58
Cognitive Robots Balancing Act including Emotions
59
Any Questions?