A Cognitive Approach to Robot Self-Consciousness Antonio Chella and Salvatore Gaglio University of Palermo Consciousness and Artificial Intelligence: Theoretical foundations and current approaches AAAI Symposium, Washington DC, 8-11 November 2007
• One of the major problems towards effective robotic architectures is to give a robot the capabilities of introspection, i.e., to reflect about itself, its own perceptions and actions during its operating life. • The robot introspection grows up from the content of the agent perceptions, recalls, actions, reflections and so on in a coherent life long narrative.
McCarthy • • What am I doing? What does my belief in p depend on? • • What is my goal? What are my choices for action? • • Observe physical body Can I achieve possible-goal? • • Do I know that proposition? Does my mental state up to now have property p? • Do I know what thing is? • • How can I plan my thinking on this problem? • Did I ever do action? When and precisely what? • What is currently happening? • What is the state of the actions I am currently performing? • What are my intentions? • Did I ever do action? When and precisely what?
Sloman
• We propose a model of robot introspection based on higher order perceptions of the robot during time. • First order robot perceptions are the immediate perceptions of the outer world of a self reflective agent. • Higher order perceptions are the perceptions during time of the inner world of the agent. • Higher order perception are at the basis of the introspctive reasoning of the robot
The Cognitive Architecture Chella, A.; Frixione, M.; and Gaglio, S. (1997). A cognitive architecture for artificial vision. Artificial Intelligence 89:73–111.
The Subconceptual Area • Low level processing of data coming from sensors. • Information is not yet organized in terms of conceptual structures and categories. • Extraction of the 3-D model • Kalman filters
The Conceptual Area • A Conceptual Space CS (Gärdenfors, 2000) is a metric space whose dimensions are strictly related to sensory based quantities (Color, pitch, spatial coordinates, etc.). • Dimensions do not depend on any specific linguistic description. • The conceptual primitive is a knoxel , i.e. a point in CS.
The Static CS • A knoxel is a superquadric • An object is a composition of superquadrics
The linguistic area • Hybrid formalism in the KL-ONE tradition • T erminological component • terminological language: semantic networks (SINets) • concept descriptions (general knowledge) • Assertional component • assertional language: ground atoms • information about specific scene
Terminological component
Assertional component • First order logic • Concepts → One place predicates • Roles → Two place predicates
Mapping
Generation of assertions • Is driven by the focus of attention • Implementation: artificial NNs • two modalities: • associative expectations • linguistic expectations • associative expectations are learned by NNs • Hebbian mechanism • linguistic expectations are driven by linguistic KB
Focus of attention
Linguistic expectations Cylinder-shaped(#k1) Box-shaped(#k2) Hammer (Hammer#1) has-part(Hammer#1,#k1) has-part(Hammer#1,#k2) A priori knowledge of the object shape
Associative expectations Hammer (Hammer#1) Box (Box#1) Next-to(l#1) Has-part(l#1,Hammer#1) Has-part(l#1,Box#1) Free associations among previously seen objects
System at work Cylinder-shaped(#k1) Box-shaped(#k2) Hammer (Hammer#1) has-handle(Hammer#1,#k1) has-head(Hammer#1,#k2) Ball-shaped(#k3) Ball(Ball#1) has-part(Ball#1,#k3) Ellipsoid-shaped(#k4) Mouse(Mouse#1) has-part(Mouse#1,#k4)
Dynamic scenes • Generic movements are made of smooth functions of time separated by instantaneous discontinuities (Marr). • A simple motion - delimited by two discontinuities - can be approximated by the superimposition of frequency harmonics (FFT analysis) Chella, A.; Frixione, M.; and Gaglio, S. (2000). Understanding dynamic scenes. Artificial Intelligence 123:89–132.
FFT Analysis of simple motion of a sq Recovered motion Parabolic motion FFT of the motion by the first frequencies of FFT
Static and Dynamic CS
Actions and Situations • A Situation is a configuration of knoxel in CS: objects maintain their motions states • An (instantaneous ) Action is a scattering of knoxels in CS: an event occurs, and some objects may change their motion state
Dynamic focus of attention • Synchronic attention : scan operation in the same CS frame. • Diachronic attention : Scan operation in subsequent CS frames. A scattering aoccurs.
Terminological component * part_of_CMotion * Composite Simple Simple Motion Motion part_of_Action * * Action * part_of_Stretch_out#1 Stretch out Forearm stretching 1/1 1/1 part_of_Stretch_out#2 * * * part_of_Seize#1 Arm approach Upper arm Seize stretching 1/1 * part_of_Seize#2 Grasp 1/1
System at work A seizes an object
Mapping
Forearm_stretching(k_a) Upper_arm_stretching(k_b) Stretch_out(st1) part_of_Stretch_out#1(st1,k_a) part_of_Stretch_out#2(st1,k_b) Arm_approach(aa1) Grasp(g1) Seize(s1) part_of_Seize#1(s1,aa1) part_of_Seize#2(s1,g1) Assertions generated at the linguistic level
Robot introspection • We propose that robot introspection is based on higher order perceptions. • First order perceptions: the perceptions of the outer world; they generate the agent conceptual space • Higher order perceptions: higher-order knoxels.
Second order perceptions
First order perception Second order perception
Second order knoxel • A second order knoxel at time t now describes the perception of the conceptual space of the agent at time t-d. • The agent perceives itself and its environment
Ax(0) Ax(1)Ax(2) Ax(0) Ax(1)Ax(2) Ax(3) Ax(3) kb kb K a K a A y (0) A y (0) A y (1) A y (1) A y (2) A y (2) A y (3) A y (3) A z (0) A z (0) A z (1) A z (1) A z (2)A z (3) A z (2)A z (3) Ax(0) Ax(1)Ax(2) Ax(0) Ax(1)Ax(2) Ax(3) Ax(3) ka ka kb kb k'a A y (0) A y (0) A y (1) A y (1) A y (2) A y (2) A y (3) A y (3) t - δ t - δ A z (0) A z (0) A z (1) A z (1) A z (2)A z (3) A z (2)A z (3) Sychronic Diachronic
Higher-order perceptions
Higher-order perceptions • The outlined procedure may be generalized to consider higher order knoxels • Higher order knoxels correspond to the robot's higher order perceptions of the knoxels of lower order at previous d times. • The union of first-order, second-order and higher- order knoxels is at the basis of the robot introspection. • The robot recursively embeds higher-order models of its own CS's during its operating life.
Higher order knoxels
Higher order CSs • Higher-order knoxels are mapped to meta- predicates in the linguistic area, i.e. predicates describing the robot perceiving itself and its own actions. • These meta-predicates form the basis of the introspective reasoning of the robot
Introspection • The forms of introspection that are more directly related to perceptual information can take great advantage from the proposed representation in the conceptual area. • In the proposed framework the current and past situations and actions, the goals, the plans the can be simply analyzed by geometric inspections in the CS • The more “abstract” forms of introspection, that are less perceptually constrained, are likely to be performed mainly within the linguistic area.
Conclusions • Open problems: • 3D real time representation of the perceived scene • storing at time t the information of higher order conceptual spaces at previous times, starting from the beginning of the robot life
Thank you !
Recommend
More recommend
