from intelligent control to cognitive control: a perspective from cognitive robot engineering point...

36
From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems Vanderbilt University

Upload: kevin-rodriguez

Post on 28-Mar-2015

216 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

From Intelligent Control to Cognitive Control:A Perspective from Cognitive Robot Engineering Point of View

Kaz Kawamura

Center for Intelligent SystemsVanderbilt University

Page 2: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Background

Our group have been working on a robotic system called ISAC (Intelligent Soft Arm Control) since late 1980s (as an industry-sponsored project.)

ISAC was initially developed as a robotic aid system using vision, voice and haptic-based adaptive control.

Page 3: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Background

Our group have been working on a robotic system called ISAC (Intelligent Soft Arm Control) since late 1980s (as an industry-sponsored project.)

Long-term goal was to develop an assembly “horon” (i.e. a cognitive co-worker) for intelligent manufacturing systems.

Page 4: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Background

Our group have been working on a robotic system called ISAC (Intelligent Soft Arm Control) for the last fifteen years.

ISAC was initially developed as a robotic aid system using vision, voice and haptic-based adaptive control.

Over the years, we gradually added hardware components and adopted a modular software development approach, i.e. multi-agent-based “hybrid architecture ( more like one Troy Kelly mentioned)”.

Page 5: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Background Our group have been working on a humanoid robotic system

called ISAC (Intelligent Soft Arm Control) for the last ten years. ISAC was initially developed as a robotic aid system using haptic-

based adaptive control.

In the last several years, we are adding computational modules to incorporate some of cognitive psychology (i.e. an central executive (A. Baddeley)) and neuroscience (i.e. an adaptive working memory (David Noelle))-based models to realize “cognitive control “ functionalities to ISAC.

Page 6: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Are these robots intelligent, cognitive or neither?

COG, MIT (Is COG the “Father of cognitive robots”?) ISAC, Vanderbilt Robonaut, NASA (Is it a vision of an ultimate

cognitive robot?) Many others shown by the workshop

participants (Rolf, Olaf, Owen, etc.)

Page 7: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Hypothesis

Artificial cognitive agents must share key features and “neurobiological and cognitive principles” (Jeff Krichmar) with humans if they are to become effective partners and coworkers in the human society.

Page 8: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Process of Cognitive (or Executive) Control

Human (and some animal) brain is known to process a variety of stimuli in parallel and choose appropriate action under conflicting goals. (Figure below was taken from: P. Haikonen, The Cognitive Approach to Conscious Machine, 2003)

Page 9: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Human Cognitive Control Functions

Ability of the brain to execute task and resolve conflicts

Focus on task context and ignore distraction Involves action selection and control where

reactive sensorimotor-based action execution falls short of task demands. Example: Stroop test

Modified from: Miller, E.K., Cognitive Control: Understanding the brain’s executive, in Fundamentals of the Brain and Mind, Lecture 8, June 11-13, 2003, MIT.

Cognitive Control

Vanderbilt University

Page 10: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

NASA-JSC Robonaut Demo:“An Ultimate Cognitive

Robot?”

Page 11: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Key Features of Cognitive Robots(A Partial/Unproven/Controvertial List )

Ability to perceive the world in a similar way to humans (or better) (e.g., “active perception”, Olaf Sporns, “ecological approach to perception”, JJ Gibson)

Ability to develop cognition through sensorymotor coordination (e.g., “morphological computation”, Rolf Pfeifer)

Ability to communicate with humans using natural language and mental models (robust HRI such as overcoming the frame of reference problem, Alan Schultz)

Ability to have a sense of self awareness (internal model and machine consciousness, Igor Alexander, Owen Holland vs. Kevin O”Reagan)

Ability to use attention and emotion to control behaviors (cognitive control)

NASA’s Robonaut

Vanderbilt University

Page 12: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Concept of a Cognitive Robotic System

Adapted from a DARPA ITPO Program web site, 2003.

Page 13: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Working Definition

Cognitive Control for robots is the attention- and emotion-based robust sensory-motor intelligence to execute the task in hand or switch tasks under conflicting goals.

Page 14: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Action

Stimuli

Actuators

Sensors

Behavior 1 …Behavior N

Behaviors

LegendSES= Sensory EgoSpherePM= Procedural MemorySM=Semantic MemoryEM=Episodic MemoryCEA=Central Executive Agent

STMAttentionNetwork

SES

SM EM

LTM

PM

Self Agent

CEA

HumanAgent

Atomic Agents

PerceptionEncodings

Head Agent

Hand Agents

Arm Agents

WorkingMemorySystem

Completed

Currently being implemented

Page 15: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Cognitive Control on ISAC

Ability to use attention and emotion to control behaviors (i.e., cognitive control) is being implemented using the Sensory EgoSphere, the Attention Network, Emotion, the Working Memory System, the Central Executive Agent, and others.

Vanderbilt University

(chunks)

SES

to WMS

Stimuli

AttentionNetwork(Gating)

PerceptionEncodings

EmotionalSalience

Taskcommand

Page 16: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Current Work Current Work is aimed at testing

how modules involved in cognitive control work together as a system:

1. Working Memory System Training

[Poster Presentation by Stephen Gordon]

2. Situation-based Action Selection

Page 17: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

1. Control Structure used during working memory system training

Page 18: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Experiment I: Working Memory Training for a Percept-Action Task

1. ISAC is trained to recognize specific objects

i.e., several colored bean bags.

2. ISAC is taught a small set of motion behaviors

i.e., reach, wave, handshake.

3. Bean bags are rearranged.

4. ISAC is asked to “reach to the bean bag”

(color is not specified).

Vanderbilt University

Page 19: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Experiment I

1. ISAC is trained to recognize specific objects ,i.e., several colored bean bags.

2. ISAC is taught a small set of motion behaviors ,i.e., reach, wave, handshake.

3. Bean bags are rearranged.

4. ISAC is asked to “reach to the bean bag” (color is not specified).

5. ISAC will attempt to load the relevant “chunks” into WMS for appropriate:

action to take (reach, wave, etc.) percept to act upon.

6. Over time, ISAC should learn which “chunk” (i.e., a percept-behavior

combination) is the most appropriate to choose

Vanderbilt University

Page 20: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Working Memory System Training

LTM

SES WM

Memory chunks

Candidate Chunks List

.

.

.

.

LearnedNetworkWeights

Percepts

Page 21: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Experiment I (cont’d)

Sample configuration for reaching(top view)

Second sample configuration(top view)

Vanderbilt University

Page 22: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Experiment I - Video

Vanderbilt University

Page 23: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Learning Results for Reaching Action

Page 24: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Experiment II: Situation-Based Task Switching (Under Investigation)

Vanderbilt University

Music - pleasure

Alarm - annoyance

perform task?dance?

alarmed?ignore?

Barney - Task

Page 25: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Experiment II

A simulation experiment to test key system components for cognitive control using CEA, attention network, and emotion

A simple situation-based task switching using the Focus of Attention (next slide) is being

Music - pleasure

Alarm - annoyance

perform task?dance?

alarmed?ignore?

Barney - Task

Focus of Attention

Percept A

Percept B

Focus of Attention

Percept A

Percept B

Focus of Attention

Percept A

Percept C

Situation S2

Situation S1

Perceptualevent

Perceptualevent

Page 26: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Situation-based Action Selection (Under investigation)

Action A1

Action A2 Selected action(execution phase)

Updateprobabilities

Appropriate action

provided by human teacher(teaching phase)

][ )(ijAP

Situation Si

P1

P2

P1

P2

P1

P2

P3

P1

P2

P3

FOA FOA

P1,P2,P3 = percepts

Perceptualevent

Page 27: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Experiment II - Video

Vanderbilt University

Page 28: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Simulation Results

0

1

2

3

4

5

6

7

8

9

10

0 10 20 30 40 50 60 70 80 90 100

Trial

# C

orr

ect

sele

ctio

ns

per

10

tria

ls

Page 29: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

What have we learned so far? Effectiveness of using a computational

neuroscience-based working memory model for perception-behavior learning on a robot (proof of concept)

Computational time of the WM software library is expected to grow exponentially as the robot accumulates experience (classical AI problem) (effective use of episodic memory?)

WM model does not seem effective for task switching

Needs a better mechanism than a FOA-based situational change for task switching (=> dynamic modeling of situations)

Page 30: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

For further information, please visit our website at:http://eecs.vanderbilt.edu/CIS/

Vanderbilt University

Page 31: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Background

Our group have been working on a humanoid robotic system called ISAC (Intelligent Soft Arm Control) for the last ten years.

ISAC was initially developed as a robotic aid system using sensor-based intelligent control.

Page 32: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Human Agent

Page 33: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Self Agent

The key agent in our cognitive architecture is the Self Agent.

Minsky calles it the “Self Model” in his forthcoming book, The Emotion Machine.

Actually he uses the term “Self Models” which include both the Self Agent and the Human Agent in our architecture.

Page 34: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Behavior 1 …Behavior N

Behaviors

SESSM

EM

PM

Self Agent

STM LTM

HumanAgent

LegendSES= Sensory EgoSpherePM= Procedural MemorySM=Semantic MemoryEM=Episodic MemoryCEA=Central Executive Agent

Central ExecutiveAgent

DescriptionAgent

Anomaly DetectionAgent

Mental ExperimentAgent

Intention Agent

Activator Agent

Emotion Agent

AtomicAgents

First-orderResponse Agent

Completed

Currently being implemented

Not yet implemented

WorkingMemorySystem

Page 35: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Central Executive Agent (CEA):Robotic Frontal Lobes responsible for cognitive control functions

Inspired by the “central executive” from Baddeley’s working memory model (Baddeley, 1986)

Functions of CEA include Obtaining task sequence for task execution

Decision making

Action execution

Task monitoring

A. Baddeley, Working Memory, 11, Oxford Psychology Series, Oxford: Clarendon Press, 1986.

DecisionMaking

TaskExecution

Task-relatedPercepts

ResponseTo Percepts

FromInitial

Knowledge

FromEnvironment

Task executionsequences

Candidate TaskExecution Sequences

Selected TaskExecution Sequences

Action

Feedback

Vanderbilt University

Page 36: From Intelligent Control to Cognitive Control: A Perspective from Cognitive Robot Engineering Point of View Kaz Kawamura Center for Intelligent Systems

Questions

1. How could cognitive control be implemented in robotics? (model or no model?)

2. How does one know when a robot becomes a cognitive robot?

Vanderbilt University