650 likes | 765 Views
3차원 안구모델의 적응적 제어. 서울대학교 뇌과학협동과정 2002-20618. Introduction. Robotic eye 연구의 의의와 현황 최근 휴머노이드 및 로봇의 연구가 활성화되고 있슴 로봇에 장착될 인공안구의 필요성 역시 대두됨. 자연스러운 안구의 움직임은 의사 및 감정의 소통에 매우 중요함. 양질의 시각정보 제공을 위해선 빠르고 정확한 안구운동이 필요. 소형화 및 경량화 요구. Eye robots. ATR 연구소의 Infanoid. ATR 연구소의 DB.
E N D
3차원 안구모델의 적응적 제어 서울대학교 뇌과학협동과정 2002-20618
Introduction • Robotic eye 연구의 의의와 현황 • 최근 휴머노이드 및 로봇의 연구가 활성화되고 있슴 • 로봇에 장착될 인공안구의 필요성 역시 대두됨. • 자연스러운 안구의 움직임은 의사 및 감정의 소통에 매우 중요함. • 양질의 시각정보 제공을 위해선 빠르고 정확한 안구운동이 필요. • 소형화 및 경량화 요구.
Eye robots ATR연구소의 Infanoid ATR연구소의 DB KIST의 HECtor MIT의 Kismet MIT의 Cog
Social role of eye-movement Communication using eye movement : pointing, mimic, social gesture Human-like Impression
y x z 3 DOF manipulator Human eye Linear motor 2dof eye
CCD camera Image input Encoder input D/O board Image capture board Motor 1 LM629 LMD18200 LM629 Motor 2 LMD18200 D/I board LM629 computer LMD18200 Motor 3 motor input Hard ware scheme
Problems Calibration : initial point fixation Motor control Velocity modification
X Z k i j Y Mechanics of 3D eye X: 시선방향 Y: 좌우방향축 Z: 상하방향축
Mechanics of 3D eye 2 skew symmetric matrix 회전축 w 를 중심으로 θ 만큼 회전
Motor property Motor 1,2,3 position relationship Motor equation :
Target board Test board
Detect salient target |(R-G)| + |(R-B)|
Psychological inspired neural net Stress release, emotion Memory : LTM, STM and Emotion Hebb’s Rule
Stress Stress : be arisen from sensory stimulus & pain. Can be released to actuator Informational energy Drive force of action Motivate movement Movement uses the Stress as Energy
Emotion Happy : Maintain this situation Unhappy(Pain) : Change this situation Happy : reducing Stress Pain : Increasing Stress Emotion makes Memory : Strengthen the links of activate cells in pool
Data derivation Data Interaction Motor cortex World Neo-cortex Making stress & Data transfer Thalamus Sensory Cortex Hippocampus Amygdala Data compression Make Memory Emotion evaluator Brain metaphor
Sensory Cortex Neo cortex Stress generator Input cells Emotion generator memory Motor Cortex Psychological inspired neural net Output cells
Animal analogy food Range of smell
LTM & STM model W LTM time STM
Memory Sensory cortex
Memory Sensory cortex Activate!
Memory Sensory cortex Activate! Activation decay
Memory Sensory cortex Activation inhibited
Memory Emotion! Sensory cortex
Memory STM duration All Links strengthened!
Memory Motor Cortex
Bnew = W * Bold W new = W old(activation > threshold) * T(1.2) M = W * (Bold + Snew) B = S + I + M S = sensor cell I = inter cell M = motor cell W = weight matrix
target Step by step movement Problem space
Learning Process Learning! New target Random search Known situation Learned action Associative memory Association & learning
Trial-and-error learning Mnew = Mold * Vstep * R(error) M = motor value Vstep = modifying size(5 ~ 15%) R = probability function(20~80%) Accept = 0.5 + D(distance)
Result Error_before – Error_after Error_before 47% = * 100
Before learning After learning Rotation value of 3 motors
Weight change of Neural net before after
Discussion Initial point fixation Effect of gravity Saccadic suppression Circular CCD
Psychological inspired neural net vs traditional neural net Emotion evokes memory STM Stress Auto weight decay Time serial associative memory Run & Learn Mixed layers
Application Navigation robot’s learning rule Motion correcting of robot Interactive controller
Conclusion • For spherical parallel 3D eye model : • Initial point fixation using visual input • Modify acute motor value by trial-error learning • Step-by-step movement by psychological inspired neural networks • Error reduced about 53% less than before learning.