Cognitive and Developmental robotics: Modelling sensorimotor, cognitive and social processes in robots and humans
Pierre-Yves Oudeyer Project-Team INRIA-ENSTA-ParisTech FLOWERS
http://www.pyoudeyer.com https://flowers.inria.fr
Who I am Organization of the course 11 jan. : Introduction to Developmental and Cognitive Robotics, The role of the body in cognition (PY Oudeyer)
25 jan. : Models of developmental learning processes in robots and humans (PY Oudeyer)
1st Feb: Models of language evolution, The open-source Poppy platform (PY Oudeyer)
8th feb. : Machine learning applied to robotics (M Lopes)
7th march : Robot learning by demonstration/imitation (M Lopes)
14th march : Human-robot interaction, user-centered design, assistive robotics
22nd feb., 29th feb., 21st march: TD M. Quinson: (1) Discovery and programming of the Poppy Humanoid robot, (2,3) Practicing algorithms for autonomous exploration and learning of sensorimotor models; (4) Stochastic optimization applied to learning motor primitives; (5) Learning by demonstation of complex gestures; (6) Design of expressive motions
A bit of history Fascinating spatio-temporal structures in the biological world
Body shapes Collective constructions
Internal body structural modularity
Motor skills Cognitive representations, Abstract social cognitive constructions Three interacting time scales for the generation and selection/ learning of new structures Self-organization in complex systems
Self-organization: formation at a macro-level of shapes/structures/ (as-)symmetries based on low-level physical laws which do not encode explicitly a map of these structures
è Typical of complex dynamical systems Rayleigh-Bénard cells Chaos
Non-linearity
Complexity increases
Broken symmetry uniform
(Adapted from Tritton, 1988; and Velarde) Development: a complex dynamical system with several spatiotemporal scales
è Need for a systemic approach … and a precise scientific language for expressing and testing theories Mathematical models
D’arcy Thompson On Growth and Form (1912) Computational models
The Chemical Basis of Morphogenesis (1952) Alan Turing
Robotic models
The Electric Dog Hammond and Miessner (1912)
Elmer and Elsie Walter (1953) Tools: computational and robotic models and experiments
Explore the landscape of complex-system mechanisms, to enhance our intuitions
è Stimulate reflexion by exploring new hypothesis spaces and verify the coherence of existing hypothesis
è Embryology (ex. Turing and morphogenesis, (Turing, 1952), Complex evolutionary dynamics (e.g. Maynard Smith and theoretical biology, Maynard-Smith, 1968), ethology of insect societies (e.g. Camazine et al., 2001);
Behavioural and Cognitive Development in Human Infants
• How do developmental structures form? • How do developmental structures impact the acquisition of novel skills? Human cognitive development
Cognitive development in robots
Developmental robotics Vaucanson, 18th century
Automate Karakuri, 19th century Automatons Robo Industrialus
Unimate, 1954, 1960 Unimate, 1954, 1960
8 AC 1946, ACE Unimate, 1954, 1960 Turing
8 AC 1946, ACE Turing
8 AC Robo Exploratius
1946, ACE Turing 20’s-30’s AC 8 AC Robo Mobilis
1946, ACE Turing 40’s AC 20’s-30’s AC 8 AC Robo Domesticus Servicius
1946, 60’s ACE AC Turing 40’s AC 20’s-30’s AC 8 AC Robo Domesticus Socialus ???
70’ s 1946, 60’s AC ACE AC ??? Turing 40’s AC 20’s-30’s AC 8 AC
Human cognitive development
Cognitive development in robots
Developmental robotics Intrinsic motivation, active learning • Autonomous collection of data Families of • Efficient learning • Self-organization of developmental trajectories developmental Social learning, imitation « forces » • Imitation of trajectories and goals • Learning combinatorial motor primitives • Optimal teaching
Cognitive abstractions: • Perceptual categories grounded in action • Active goal babbling, macro-actions, macro-states • Efficient learning in high-dimensions
Body morphology and growth : • Morphology • Synergies • Self-organization of movement structures • Adaptive maturation driven by intrinsic motivation • Self-organization of maturational schedule
1. What is walking/locomotion?
Calculating optimal commands? From complex planning of walking steps on (lab controlled) uneven terrain …
Requires high-quality sensors, models of own’s geometry and dynamics, complex planning computation achieved at a very high frequency (around 1kHz) à Yet slow and fragile movements ! Walking faster than the brain can think …
Movements generated directly by materials interacting with the laws of physics à Compliance, elasticity, distribution of weight à Morphology generates movement directly, no neural “computation” à “Morphological” computation … transposed to robots From planning (fragile and costly) walking patterns …
… and fighting against gravity … … to spontaneous physically generated walking patterns, leveraging gravity for energetically cheap movement
Tad McGeer (McGeer, 1990), Nagoya Univ. (2005) Morphological computa on
• Collaboration with Labri/ Univ. Bordeaux I
• Collaboration with J-R. Cazalets, Integrative Neuroscience Institute, Bordeaux (Ceccato et Cazalets, 2009)
The Acroban humanoid (Ly, Lapeyre, Oudeyer, 2011, IROS)
Shape of leg thighs
Experiments about the impact of We measured a reduction by 50% the of thigh shape over the walking upper body perturbations during dynamic walking (Lapeyre et al., 2013)