380 likes | 588 Views
Developmental Robotics: and the importance of Psychology. Mark Lee mhl@aber.ac.uk. Aberystwyth, Wales. Acknowledgements. Martin H ülse , Tao Geng , Mike Sheldon, James Wilson, Patricia Shaw, James Law, Sebastian McBride Fei Chao, Qinggang Meng , Tom Izzett …
E N D
Developmental Robotics: and the importance of Psychology Mark Lee mhl@aber.ac.uk
Acknowledgements • Martin Hülse, Tao Geng, Mike Sheldon, James Wilson, Patricia Shaw, James Law, Sebastian McBride • Fei Chao, QinggangMeng, Tom Izzett … • And partners in Reverb, Rossi, Im-CleVeR … Funding: EPSRC (DVL & Reverb) EC FP7 (Rossi & ImCleVr)
Advanced Robotics: Key requirements • Adaptive behaviour • Cumulative learning from experience • Autonomous handling of new situations • Growth of competence • Key words: Autonomy & Growth
Premises • No really intelligent robots yet exist • Classical AI has been insufficient • Development is essential for human learning • Development may be essential for robot learning if we wish to capture similar performance
Biologically Inspired Robotics • Anatomical- structural and process models of systems • Evolutional- growth and change within a species • Developmental - adaptive growth and change in the individual
Inspiration from Alan Turing “Instead of trying to produce a programme to simulate the adult mind, why not rather try to produce one which simulates the child's? If this were then subjected to an appropriate course of education, one would obtain the adult brain [...]” A.M. Turing, Mind, 59, 433-460, 1950
Key issues and observations • Stages (in behaviour) • Constraints – many sources • Shaping – by constraint & interaction • Cumulative learning • Intrinsic motivation • Intrinsic activity (motor babbling) • Goal-free or goal-driven
Motor babbling and Play • Motor activity: • - produces inputs from sensors • - provokes the environment • gives new viewpoints • Infant play - Essential behaviour
Motor babbling in infants • sucking • eye movements • head rolling • facial expressions • body and limb kicking actions • reaching • touching
Our approach • Use salience array for stimuli selection • Motor babbling when quiescent • When actions correlate with stimulus, try to repeat exact action • Noise and natural variation will test extent of new stimulus/action correlation. • Those correlations that withstand repeated action can be recorded in mappings, associations, or other means.
Our framework Computational substrate: sensory-motor mappings Bidirectional linking of SM maps, learned through Hebbian correspondence Produces egocentric SM space
Learning sensor-motor mappings Learning saccadic eye movements retina space relative gaze space With Fei Chao
Learning sensor-motor mappings Hand-eye coordination abs. gaze space reach space 2.0 +/- 1.0 cm With Martin Hülse, Sebastian McBride
Infant (to iCub) development Survey and study of literature on infant development Constructed a timeline of activity from conception to 12 months Prepared similar development chart for iCub Constructed constraint network and identified dependencies
Infant vision development • Increase in image resolution (birth to 12 months) • Widening field of view (6-10 weeks, 20-40 degrees) • Increased sensitivity to stimulus (birth to 6 months) • Increased Image transfer rate (birth to 3 months) • Increased focal range (1 to 2 months), initially ~21cm • Increased colour resolution (birth to 4 months) • Stereopsis onset and improvement (3 to 12 months) • Migration of rods and cones – At birth the distribution of rods and cones is roughly uniform, with migration to adult positions occurring over the first 11 years. Fastest migration occurs 2-3 months after birth.
Integrated architecture retina Saccadic eye movements gaze
Integrated architecture head Head and eye movements Hand-eye coordination gaze
Integrated architecture retina gaze Arm movements Hand regard Hand-eye coordination reach
Integrated architecture retina gaze Hand movement Grasp control reach grasp
Integrated architecture retina two arm coordination visual memory gaze reach reach grasp
Stages seen infollowing video 1 – early eye saccade learning 2 – later part of above stage 3 – early head movement learning 4 – better head control at later part of above stage 5 – hand regard behaviour (combining gaze and reach spaces) 6 – Early reaching (plus hand regard) 7 – Reaching with stimulus lights included 8 – more accurate reaching?
iCub development Initial learning trials • Eye saccades 4 minutes 25 fixations • Head rotation 9 minutes 45 fixations • Arm reaching 10 minutes 22 locations
Consequences Play and babbling – notrandom behaviour, nota side-effect of learning Goals can be created – not given (where would they come from?) Solitary and socialshaping - both vital Akey principle for autonomous development?
Advantages of this Approach • Intrinsic activity: significant principle • Autonomous, generates experience • Unsupervised,incremental, continuous and cumulative learning • Self calibration • Shaping: main interaction mode • Efficient: very fast learning/adapting
The Future ? Two types of robot: • Developmental • Task based
The Future ? • No programming - only training (in the world) • Customised in situ. • Experience resides in systems, but all different, (sets of individuals). service centre = robot remedial school, programs of corrective shaping. But we can also examine brain! so transferable skills?