CSG6120-01인지과학특강 I인지과학특강 I
(Theory of Mind for Robots)
Spring Semester, 2011
Teaching Staff Lecturer
- Sung-Bae Cho (C515; 2123-2720; [email protected])- Kyung-Joong Kim ( 3408-3838, [email protected])y g g ( , j@ j g )
Hours : 화요일오후 2~4시, 오후 5시~6시
Place : A542
Office Hours : 화요일 4시~5시
Robot and Cognitive ScienceCognitive RoboticsRobot Psychology
Experimental Robotics Developmental Robotics
Bio-Inspired Robotics
CognitiveScience
Robotics
Cross Fertilization
• Biological Inspiration• Computational Model
CognitiveScience
Robotics
• Engineering Inspiration• Engineering Inspiration• Virtual Humans and Animals
• Reverse Engineering Cognitive Functions
Theory of Mind
Theory of mind is the ability to attribute mental states—beliefs, intents,
desires pretending knowledge etc to oneself and others and todesires, pretending, knowledge, etc.—to oneself and others and to
understand that others have beliefs, desires and intentions that are
different from one's own
from wikipedia
Course Goals
Theory of Mind에대한이해및최신연구동향파악
인지과학의연구도구로Robotics 기술이해
Robot을위한 Theory of Mind 관련연구이해및연구주제수행 y
Course Materials
Reading in Theory of Mind for Robots– Theoretical Foundations for Theory of Mind
• Theory, Neuroscience, Animal Study ….
– Applications for Simulated Robotspp• Virtual Character, Robot Simulator
A li ti f R l R b t– Applications for Real-Robots • Real Robots (E-Pucks, Humanoids, …)
Week 2시간 1시간
논문발표 (학생)
Week 2시간 1시간1 Course Introduction
2E-Puck 로봇소개
E Puck Simulator소개 신경망 진화연산소개2 E-Puck Simulator 소개IR 센서소개
신경망, 진화연산소개
3 PhysX 소개OptiTrack 소개 진화신경망소개OptiTrack 소개
4 ToM 기초논문들 Simulated Robot을위한 ToM
5 ToM을로봇에적용한대표사례논문들
Real Robot을위한 ToM논문들
6 EEA 알고리즘및그응용 제공 SW 소개7 Term Project Proposal8 Midterm Exam8 Midterm Exam9 로봇을위한 ToM 프로젝트진행상황 Discussion
10 석가탄신일 석가탄신일11 로봇을위한 ToM 프로젝트진행상황 Discussion11 로봇을위한 ToM 프로젝트진행상황 Discussion12 로봇을위한 ToM 프로젝트진행상황 Discussion13 로봇을위한 ToM 프로젝트진행상황 Discussion14 로봇을위한 ToM 프로젝트진행상황 Discussion14 로봇을위한 ToM 프로젝트진행상황 Discussion15 Term Project Final Presentation16 Final Exam
Resources
E-Pucks Optical Tracking Devices (OptiTrack) IR Light Source (SiMa Night Vision)
Evaluation Criteria
Evaluation Criteria– Term Project (written report and an oral presentation) : 70%– Class Presentation : 20%– Class Participation : 10%
Term Project (Oral presentation is required) :– Theoretical Issue (Analysis, Experiment, Simulation) : Originality
I t ti P i (G D t ) P f– Interesting Programming (Game, Demo, etc) : Performance– Survey : Completeness
Basics in Theory of MindBasics in Theory of Mind
Mind Reading
Non-Human Theory of MindNon Human Theory of MindThey Can Do?
Theory of Mind Cartoons
Non Theory of Mind Cartoons
Recursive Nature of ToM
Robotics Approach(Self Modeling)
Robotics Approach(Theory of Mind)
Theory of Mind for Humanoids [MIT, 2000]
GazingFace DetectionFace Detection Infer Simple Intention
[Brian Scassellati][ ]
Mario et al. (2007)
Robotics Approach(Neural-basis)
Theory of Mind in Robotic Society
A: I Know “You”! B: I Know “You Know Me”!A I K “Y K I K Y ”!A: I Know “You Know I Know You”!B: …
Predator Pray
Scenario
1 Infer Other’s Brain 2 Exploitation
From Virtual To Real
1 PhysX 2 EnKi 3 E-Puck
Controller (Neural Network)
b1b4w7
Steering AngleLeft Light Sensor Value w1w2
b2
bw7
w8w3
w4
b2
b5
w9
w10VelocityRight Light Sensor Value
w4w5
w6b3
b5w11w10
w6w12
-0.509588
-0.385503
-0.104083
0 172043
0.238755
0.172043
0.272120
0 1922550.192255
-0.973429
9.644485
66.393120
[Original] [Reconstructed]speed[Original] [ eco s uc ed]
Sens
or V
alue
Sens
or V
alue
Lef
t Lig
ht S
Lef
t Lig
ht S
Ri ht Li ht S V l Right Light Sensor ValueRight Light Sensor Value Right Light Sensor Value
Real Robot Scenarios
Possible ProjectsPossible Projects
Overview
Robot
B h i
ToM
Behavior (Trajectories)
인공신경망(Robot’s Mind)
알고리즘
EstimatedOne
Possible Project Lists I (Survey)
Theory of Mind를설명하는이론들에대한조사
Brain-Imaging을통해밝혀진 Theory of Mind 관련사실들조사
인간이아닌다른동물들의 Theory of Mind에대한조사 y
Theory of Mind를테스트하는방법에대한조사
봇시 레이션Possible Project II (로봇시뮬레이션) ToM알고리즘의종료조건테스트 (다양한조건들중에어떤것이최적인 ToM알고리즘의종료조건테스트 (다양한조건들중에어떤것이최적인지실험적으로테스트)
T M알고리즘에서상대방로봇의행동데이터들을어떻게활용하는것이 ToM알고리즘에서상대방로봇의행동데이터들을어떻게활용하는것이최적인지테스트 (가장최근에얻어낸데이터만이용/과거데이터도모두이용/가중치를부여하여이용)
ToM알고리즘에서상대방로봇의행동데이터를어떻게얻는것이최적인지테스트 (random하게얻어온다/능동적으로필요한데이터를얻어온다)다)
ToM알고리즘에서두 Trajectory가닮았다는것을어떻게측정하는것이j y좋은가? (Euclidean Distance, Derivatives …)
ToM알고리즘에서상대방로봇의행동데이터를사용할때그 trajectory ToM알고리즘에서상대방로봇의행동데이터를사용할때그 trajectory의각 segment가서로다른중요도를가진다고보고, 그중요도를최적화한다.
실제 봇Possible Project III (실제로봇)
Reality Gap (Simulator와실제로봇) 최소화– OptiTrack이용– Sensor Models (서로다른정밀도지님) – Motor Adjustment Parameters – Reality Gap이얼마나줄어드는지평가y p
Immature Neural Controller 진화단순히직진으로빛을따라가는최적화한제어기보다는– 단순히직진으로빛을따라가는최적화한제어기보다는
– 회전하거나커브를그리는형태의제어기를선호– Long-Term Exposure Shot으로로봇행동촬영하기