lecture 3. cognitive robotrit.kaist.ac.kr/home/ee788ln?action=attachfile&do=... · lecture...

EE788 Robot Cognition and Planning, Prof. J.-H. Kim

Lecture 3. Cognitive Robot: Autonomous Computationally Intelligent System

Lecture 3. Cognitive Robot: Autonomous Computationally Intelligent System

Robot Intelligence Technology Lab.

Slides for this course are from the tutorial of Professor John Taylor, King’s College London, at CEC06.

Lecture Objectives

n WHAT?l Autonomy = ability to move validly under own steam

(possessed by many animals)l Autonomy as evolutionary treel Possession of autonomy ≠ intelligence

Eg, persistent criminals, schizophrenics (schizophrenia (mental

2Robot Intelligence Technology Lab.

Eg, persistent criminals, schizophrenics (schizophrenia (mental illness) patients), ….

l Add intelligence (computationally created)But dangerous => wars, mayhem, suicide bombers, etc, etc.

Lecture Objectives

n HOW?l Lessons from existing solution – humansl Brain arguably subtlest ‘machine’ in universe (is conscious,

not possessed by universe – insanely produces black holes)l Need emotional value/empathy (‘love’?)

=> Need careful guidance of development of ‘conscious’ machine with ‘guided creative fun’ (learns value of NO)


with ‘guided creative fun’ (learns value of NO)l As should be in human upbringing

Lecture Objectives

=> BRAIN GUIDANCEl Look for brain principles (intelligence & emotion)l Many suggestedl Propose highest level control: attentionl Biased by emotion valuations of worldl Internal rules developed with biasing valuations


l Internal rules developed with biasing valuationsl Needs to be highly adaptive

(reinforcement & error learning schemes)l Bring these to build general cognitive architecture

Brain

n Global Level: The Human Brainl A very complex system:

1011 neurons, 1014 connectionsl Many hypercolumns & modulesl Subtle functionalityl Subtle pattern emergence

--> High level functionality


--> High level functionality

Brain

n Overview of Brain Parcellation of Function


Brain

n Current information processing systems (IPS): incredibly fast, cheap, and scalable, and can be used for a large number of problems humans find very difficult.

n BUT: they fail that even a child does without much trouble (smooth walking, recognizing faces, understanding language, showing creativity…).

7

language, showing creativity…).


Brain

n Artificial IPS have been developed to solve “artificial problems” (AP). But now, we need to start addressing “natural problems” (NP) as well.

n Natural problems appear to be easy for humans because they have been solved by the brain during the evolution. Artificial problems are difficult for brains because there was no necessity for the brain to find solutions for them.

8

was no necessity for the brain to find solutions for them.

n Hypothesis: solutions to natural problems are realized as structure and function of the brain

à i.e. as Brain’s information processing architecture.


9Robot Intelligence Technology Lab.By courtesy of Dr. Daeshik Kim

What cortical stream: area V4 of the visual cortex and temporal cortexWhere processing stream: middle temporal area (MT) of the visual cortex and parietal cortex


Nature of Cognition

n Cognition defined asl “That operation of the mind by which one becomes aware of

objects of thought or perception; it includes all aspects of perceiving, thinking, and remembering”

l “Mental functions such as the ability to think, reason, and remember”

l “High level functions carried out by the human brain, including


comprehension and use of speech, visual perception and construction, calculation ability, attention (information processing), memory, and executive functions such as planning, problem-solving, and self-monitoring”

=> Cognition complex!

Nature of Cognition

n How to look at cognitionl Reasoning, planning and self-monitoring:

=> crucial components of cognitionl Leave out speech: animals can reason

=> look at nonlinguistic cognitionl Approaches for reasoning, planning and self-monitoring:

a) Symbolic: logical inference on language structures


a) Symbolic: logical inference on language structuresb) Probabilistic: cognition = probabilistic inferencec) Connectionist: how can inference be obtained from neural network structures at sub-symbolic level?

l Neural structures as most relevant to relate to brain processing methods

Problems for Autonomous Machines

n Problems of machines presently:1) Scalability (many sensors, real time)

2) Context awareness (peripheral sensitivity)

3) Robustness (against damage/loss)

4) Autonomy & self-management (stay alive)


5) User adaptability (varying profiles)

6) Fast computation (rapid guidance)

Essential Structures

n Crucial components = internal models of worldn Forward models (encode causality of the world)n Working memory modules (imagine the world)n Inverse model controller (IMC) (generates desired control

actions)n Filter control system


n Filter control system (filters world into ‘to be attended to’ + ‘rest = distracters’)

n FM/IMC by control theory (no WM or attention)

Essential Structures

n Control Model Componentsl Inverse model controller (IMC)

l Forward model/ observer


l Error Monitor Module

Solution to Autonomous Machine Problems

n Learn to Pay Attention!l Attention solves:

1) Scalability: filter out distracters2) Context awareness: attend to important3) Robustness: use neural multiplexing/Perception On Purpose4) Autonomy: determine by own goal structure


4) Autonomy: determine by own goal structure5) User adaptability: by training to a user6) Speed: by hardware nanotechnology

Solution to Autonomous Machine Problems

n Extra Components for Cognitive Powersl Show how can model reasoning powers of animals

by dedicated coupled FM/IMC/WM systemsl Used to generate and transform ‘world in head’ to achieve

executive functionsl Attention crucial in complex worldl Need good model of attention/ learning/ WM/ executive powers/


l Need good model of attention/ learning/ WM/ executive powers/ STM/ LTM: all powers possessed by humans

Cognitive Brain-based Machine Projects

n GNOSYS ‘Cognitive Robot’ (EC IST Project Oct 2004 – ‘07)

n MATHESIS ‘Learning Others’ Actions’ (EC IST Project Feb 2006-’09)

n ‘Attending to the World’ (EPSRC Project March 2004 – 2007)


(EPSRC Project March 2004 – 2007)n ‘Analyzing Attention<->Emotion’

(BBSRC Project 2005-2008)n ‘Modelling Emotion’

(EC HUMAINE NoE, 2005 - 2008)


n Focus of GNOSYS1) Develop percepts/concepts/rewarded goals/reasoning/abstraction2) Learn to perform goal-directed tasks3) Learn in novel environments4) Reasoning by forward model5) Globally integrated system6) Employ various memory types


6) Employ various memory types(STM/LTM/iconic/associative memory)

7) Interdisciplinary: Computer vision/ Cognitive Science/ Robotics/ Control/ AI/ Mathmatics


n How GNOSYS works



n GNOSYS ReasoningDomains/Environmentsl Three levels of environment

- Level 1: Learn shapes/colors; move and touch: [2]-D objects• Powers: Concept/Attn/Goals/Actions on objects/Salience of objects in environment


- Level 2: [3]-D objects and actions: pick-up, stack, learn new objects• Powers: manipulate to achieve goals

- Level 3: Hierarchy of objects; run virtual object/ action sequences to achieve goals• Powers: Reasoning/ novel objects/ actions


n The Hybrid GNOSYS Brain


HW #2: 더확장하기


n EC Cognitive Systemsl ‘Focus is on research into ways of endowing artificial systems

with high-level cognitive capabilities, typically perception, understanding, learning, knowledge representation and deliberation, thus advancing enabling technologies for scene interpretation, natural language understanding, automated reasoning and problem-solving, robotics and automation, that are relevant for dealing with complex real-world systems. It aims at


relevant for dealing with complex real-world systems. It aims at systems that develop their reasoning, planning and communication faculties through grounding in interactive and collaborative environments, which are part of, or connected to the real world.

l These systems are expected to exhibit appropriate degrees of autonomy and also to learn through "social" interaction among themselves and/or through human-agent cooperation; in a longer term perspective, research will explore models for cognitive traits such as affect, consciousness or theory of mind.’


n How To Assess?l Ambitious, even mentions consciousness and theory of mind.l Similar ambition in other new adventures in cognitive research:

Brain Sciences Institute (BSI) in Tokyo, BICA (USA) – emphasize brain basis/use of brain guidance

l BSI made good progress towards its aims l High ambition => not negative reaction


l High ambition => not negative reactionl But need careful assessment of projects & resultsl Need to realize some goals harder than initially thought

Eg, consciousness!

The EC Cognitive Systems Projects

n The 23 ACS projects (in alphabetical order):1) BACS: Bayesian Approach to Cognitive Systems2) CASBLIP: Cognitive Aid System for Blind People3) CLASS: Cognitive-Level Annotation using Latent Statistical

Structure4) COSPAL: Cognitive Systems using Perception-Action Learning5) COSY: Cognitive Systems for Cognitive Assistants


6) DECISIONS-IN-MOTION: Neural Decision-Making in Motion7) DIRAC: Detection and Identification of Rare Audio-visual Cues8) eTRIMS: eTraining for Interpreting Images of Man Made Scenes9) euCOGNITION: European Network for the Advancement of

Artificial Cognitive Systems10) GNOSYS: An Abstraction Architecture for Cognitive Agents11) HERMES: Human-Expressive Representations of Motion and

their Evaluation in Sequence


12) ICEA: Integrating Cognition, Emotion and Autonomy13) JAST: Joint-Action Science and Technology14) MACS: Multisensory Autonomous Cognitive Systems15) MATHESIS: Observational Learning in Cognitive Agents16) Mind RACES: from Reactive to Anticipatory Cognitive Embodied Systems17) PACO-PLUS: Perception, Action and Cognition through Learning of Object-

Action Complexes18) PASCAL: Pattern Analysis, Statistical Modelling and Computational Learning


19) POP: Perception On Purpose20) RASCALLI: Responsive Artificial Situated Cognitive Agents Living and

Learning on the Internet21) ROBOT-CUB: Robotic Open-architecture Technology for Cognition,

Understanding and Behaviours22) SENSOPAC: SENSOrimotor structuring of Perception and Action for

emerging Cognition23) SPARK:Spatial-temporal patterns for action-oriented perception in roving

robots


n Assessment approachl Gather the projects together under several headings, emphasizing

various approaches to cognition:– Embodiment driven (# 13, 21)– Applications-driven (# 2, 5, 7, 8, 20)– Machine-intelligence driven (# 1, 3, 18)– Neural-based (# 15)


– Neural-based (# 15)– Cognitive science based (symbolic)– Hybrid (# 10)– Dynamic systems (# 23)

l Various approaches need not be most effective to achieve breakthrough in creation of autonomous cognitive machine.

l Need general model of human cognition to properly assess projects viability.


n Other Groupsl Numerous other groups: AIBO (Sony), ASIMO (Honda),

Darwin 1–N (G Edelman, La Jolla), K Kawamura (Developmental Robotics), ATR Lab, Kyoto (M Kawato), COG Lab (R Brooks, MIT), ‘Conscious Robot’ (O Holland, Essex) + many more

l Use of HDP (BEP-based) very effective in creating motor control of walking robots (Kawato/Doya)


l ‘Imitating robots’ important areal Progress across many fronts

REFERENCES

3.1 The CODAM Model of Attention and Consciousness 3.2 Neural Networks of the Brain: Their Analysis and

Relation to Brain Images3.3 Modelling Human Attention and Emotions

n Video of robot under control of G-Brain, athttp://www.ics.forth.gr/gnosys/


http://www.ics.forth.gr/gnosys/n Results in CNS group at KCL CNS website

http://www.kcl.ac.uk/research/cns/cns.htmln Progress in modeling animal reasoningn Much more complexity needed at computational leveln Even then may need chip implementation to attain

consciousness

Cognitive Architecture Newell et al., 1982

n An artificial computational process which mimics certain acts of cognition (Langley 2006)l focuses on cognition as a whole

not some aspect of it like learning or problem solvingl The concept of modularity is central.

Mostly fixed, but some allows growth and modification over time. l usually learning capability

Ref. 3.4

35

l usually learning capabilityl Concurrency is normally required.

Three kind of processes all running in parallel: - reactive process: equivalent to things that we are doing without thinking,- deliberate process: core to reasoning and planning- reflective process: change the mental state of the robot


à Cognitive system: adaptive, anticipatory and purposive goal-directed behavior

Cognitive Architecture

n Additional features for socially interactive systems (Scassellati 2001): l acts with the same reaction time to the modeled cognitive systeml can operate in the same complex, noisy, and cluttered environment people

inhabit such that it can recognize the social cues and express itself with such cues

l Robust behavior and reliability

n Approaches:

36

n Approaches: l Cognitivist approach:

- based on symbolic information processing representational systems- Cognition as “symbolic, rational, encapsulated, structured, and algorithmic,”

l Emergent systems approach- based on principles of self-organization - Cognition as emergent, self-organizing, and dynamical

l Hybrid approach incorporating aspects of these two (Vernon 2007, Clark 2001)


COGNITIVE ARCHITECTURES reviewed in this paper

Cognitive Architecture


HW #3 하나를정해 ppt로정리하기대표논문 pdf와함께제출

[email protected]

Cognitivism

n The focus is on the aspects of cognitions, constant over time and relatively independent of the task.

n Cognitive model: The combination of a given cognitive architecture and a particular knowledge set l The cognitive architecture

- defines the manner in which a cognitive agent manages the resources at its disposal, where resources are the computational system in which the

38

its disposal, where resources are the computational system in which the physical symbol system is realized.

- specifies the formalisms for knowledge representations and the memory used to store them, the processes that act upon that knowledge, and the learning mechanisms that acquire it

n ISAC (Intelligent Soft Arm Control) - a parallel distributed cognitive control system for a humanoid robot (Kawamura 2004), Soar architecture (Rosenbloom 1993), ACT-R (Adaptive Control of Thoughts, Rational) architecture (Anderson 1996), and Society of Mind (Minsky 1986)


Emergent Approaches

n A very different view of cognition from cognitivism: Cognition is the process by which an autonomous system becomes viable and effective in its environment through a process of self-organization through which the system is continually re-constituting itself in real-time to maintain its operational identity through moderation of mutual system-environment interaction and co-determination.

- Co-determination: the cognitive agent is specified by its environment and at the same time that the cognitive process determines what is real or meaningful for the agent (Granlund 1999, Vernon 2007) .

39

- Emergent approaches are intrinsically embodied and the physical experiences of the robot play a central rule in its cognition.

- Architecture: not functionally modular, but parallel, real-time, and distributed l Assertions:

- The primary model for cognitive learning is anticipative skill construction rather than knowledge acquisition

- Processes that both guide action and improve the capacity to guide action while doing so, are taken to be the root capacity for all intelligent systems.


Emergent Approaches

n Major subcategories: connectionist, dynamical models and enactive system models l Connectionist systems with simple abstractions of the brain, rely on

parallel processing of non-symbolic distributed activation patterns using statistical properties, rather than logical rules, to process information and achieve effective behaviors (Medler 1998).

- Local aspects of the problem, rather than giving insight into how a complete system might be organized

40

system might be organized l Dynamical system theory: used to complement classical approaches

in artificial intelligence (Reiter 2001) and deployed to model natural and artificial cognitive systems (Gelder & Port 1995).

- An open dissipative nonlinear non-equilibrium system with relatively low order is used to model the system behavior.

- The ability to characterize the behavior of a high-dimensional system with a low-dimensional model, is one of the features that distinguishes dynamical systems from connectionist systems


Dynamical models

n The core is the application of the mathematical tools of dynamics to the study of cognition. l A cognitive dynamical system:

- The components of the system must be related and interact with one another.

- Any change in one component or aspect of the system must be dependent on the states of the other components: “they must be interactive and self

41

on the states of the other components: “they must be interactive and self contained.”

- “Cognition is nonsymbolic, nonrepresentational and all mental activity is emergent, situated, historical, and embodied.”

l Cognitive system: - Not a discrete sequential manipulator of static representational structures,

rather a structure of mutually and simultaneously influencing change in real time of ongoing change of environment, the body and the nervous system.


Enactive system models

n The goal: the complete treatment of the nature and emergence of autonomous, cognitive, social systems.l Based on the concept of autopoiesis – literally self-production –

whereby a system emerges as a coherent systemic entity, distinct from its environment, as a consequence of processes of self-organization.

l Enaction involves different degrees of autopoeisis and three orders of system

42

system- First-order autopoietic systems correspond to cellular entities that achieve

a physical identity through structural coupling with their environment.- Second-order systems are metacellular systems that engage in structural

coupling with their environment, this time through a nervous system that enables the association of many internal states with the different interactions in which the organism is involved and has processes of self-development..

- Third-order systems exhibit coupling between second-order (i.e., cognitive) systems, i.e., between distinct cognitive agents.


n Cognitivist : l Two central ideas: the dualism that speculates the logical separation of mind and body

and the functionalism that speculates that cognitive mechanisms are independent of the physical platform (Freeman 1999)

l The construction is top-down.l A symbolic structure is crucial for mimicking human behavior .l Independent of their physical platform, general-purpose systems of cognition operating

as a brain-in-the box l The adaptation is achieved by gaining new knowledge.

Comparison

43

l The adaptation is achieved by gaining new knowledge.

n Emergent approach: l Reflect or recognize the morphology of the physical body in which it is embedded and

of which it is an intrinsic part l It is bottom-up. l Dynamical and organizational properties of the systems e.g. a connectionist approach

is sufficient for such a task.l Embodied metaphors of their environment and the physical instantiation plays a direct

fundamental part in the cognitive process .l Adapts by changing its structure to attain new dynamics.


Comparison


Hybrid Structures

n Combine aspects of cognitivist and emergent approaches, but against the uses of explicit designer knowledge in an AI system. l Their goal is to develop an active ‘animate’ perceptual system in which

the behaviors themselves become the focus, rather than the perceptual abstractions

l The adaptability of emergent systems, but the advanced starting point of cognitivist systems

45

cognitivist systems- still use representational invariance but the difference is these representations

should only be constructed by the system itself as it interacts with and explores the world rather than through a priori specification

l Ex) combining a neural-network-based perception-action component and a symbolic component

n The Polyscheme model (Cassimatis 2002) and the Cog project (Brooks 1999, Scassellati 2001)


The Cog Project

n Methodologyl Need a theory of mind to discover the hidden properties of their environmentl Assumptions:

- Humans use physical symbol systems to “think” (Newell & Simon 1976)- The existence of a monolithic control, though not supported by neurophysiology

as there are multiple parallel channels of control dependent on the taskl Humans are not general purpose machines and they do not perform all the

tasks equally. Human intelligence also includes

46

tasks equally. Human intelligence also includes - Development: forms the framework by which humans successfully acquire

increasingly more complex skills and competencies. - Social interaction: allows humans to exploit other humans for assistance,

teaching, and knowledge. - Embodiment and physical coupling: allow humans to use the world itself

as a tool for organizing and manipulating knowledge. - Integration: allows humans to maximize the efficacy and accuracy of

complementary sensory and motor systems.


The Cog Project

n The Robotl Development

- By gradually increasing the complexity of the required task, a developmental process optimizes learning.

l Physical embodiment - For an embodied system, internal representations can be ultimately

grounded in sensory-motor interactions with the world (Lakoff 1987).

47

l Integration of multiple sensory and motor systems- The computational control is a heterogeneous network of many different

processor types operating at different levels in the control hierarchy. l Social Interaction

- Building social skills into an artificial intelligence provides not only a natural means of human-machine interaction but also a mechanism for bootstrapping more complex behavior.

- A motivational system works to maintain robot drives within homeostatic bounds and motivates the robot to learn behaviors that satiate them.


n A Theory of Mind l The ability to recognize what another person can see, the ability

to know that another person maintains a false belief, and the ability to recognize that another person likes games that differ from those that the child enjoys- Normal social interactions depend upon the recognition of other points

of view, the understanding of other mental states, and the recognition of complex non-verbal signals of attention and emotional state.

The Cog Project

48

of complex non-verbal signals of attention and emotional state. l Leslie’s Model of Theory of Mind

- The world is naturally decomposed into three classes of events: - One class for mechanical agency: The Theory of Body module (ToBY)

explained by the rules of mechanics - One for actional agency: system 1 of the Theory of Mind module (ToMM-1)

explaining events in terms of the intent and goals of agents, i.e., their actions. - One for attitudinal agency: system 2 of the Theory of Mind module (ToMM-

2) which explains events in terms of the attitudes and beliefs of agents


l Baron-Cohen’s Model of Theory of Mind - The set of precursors to a theory of mind can be decomposed into

four distinct modules - The first module, called the intentionality detector (ID), interprets

self-propelled motion of stimuli in terms of the primitive volitional mental states of goal and desire and produces dyadic representations that describe the basic movements of approach and avoidance, such as “he wants the food” or “she wants to go over there.”

- The second module, called the eye direction detector (EDD), processes

The Cog Project

49

- The second module, called the eye direction detector (EDD), processes visual stimuli that are eye-like to determine the direction of gaze, such as “agent sees me” and “agent looking-at not-me”. .

- The third module, the shared attention mechanism (SAM), takes the dyadic representations from ID and EDD and produces triadic representations of the form “John sees (I see the girl)”.

- The fourth module, the theory of mind mechanism (ToMM), provides a way of representing epistemic mental states in other agents and a mechanism for tying together our knowledge of mental states into a coherent whole as a usable theory, such as “John thinks (Elvis is alive).”


l An Embodied Theory of Mind - A hybrid architecture called the embodied theory of mind connects

modules similar to Leslie’s ToBY and Baron-Cohen’s EDD, ID, and SAM together with real perceptual processes and with links to physical behaviors and extends the current models to behavior selection, attention, and more complex behavioral forms.

- Overview of the hybrid theory of mind model

The Cog Project


lecture 3. cognitive robotrit.kaist.ac.kr/home/ee788ln?action=attachfile&do=... · lecture...

Documents