Life in the Humanoid Robotics Group MIT AI Lab Paul Fitzpatrick

Our Goal Natural tasking Respond to humanoid cues Exhibit humanoid cues

Our Methodology Integration Social Interaction Embodiment Development Development: using simple, early behaviors that serve to structure more complex behaviors. Example: the stages of gaze sensitivity and following seen in infants. Also, growing complexity of environment, best If just slightly out of reach. Social interaction: non-hostile world – contigent responses. Embodiment: grounding representations in sensory-motor interaction. Flexibility? Integration: different senses have different strengths and weaknesses but considerable overlap, good to combine them. Development

Our Robots

Research Issues Our Approaches K n o w i g h a t m e M p b d s C r c f l u z G x k I v S U y A D L P E - ( , ) Research Issues Our Approaches

Research Issues Our Approaches K n o w i g h a t m e M p b d s C r c f l u z G x k I v S U y A D L P E - ( , ) Research Issues Our Approaches

Social Interaction

Cynthia Breazeal

Showing and Reading

Evaluation Metrics Must consider human response as well as robot’s behavior Can people intuitively read and do they naturally respond to Kismet’s social cues? Does Kismet elicit social cues (scaffolding) from the human that could be used to benefit learning? Can Kismet perceive and appropriately respond to these naturally offered social cues? Does the human adapt to the robot, and robot adapt to the human, in a way that benefits the interaction? (entrainment)

Research Issues Our Approaches K n o w i g h a t m e M p b d s C r c f l u z G x k I v S U y A D L P E - ( , ) Research Issues Our Approaches

Brian Scassellati

Theory of Mind For complex learning tasks the robot needs to understand what it is the human knows about the world about the robot’s current state understand what the human is doing and why understand what is the human

Leslie’s Model ToBY Inanimate Objects Mechanical agency ToMM-1 Actional agency Divides the world into three spheres of causation, each of which is processed by a separate module: Theory of Body (ToBY) – applies naïve physics to objects Theory of Mind Mechanism (ToMM) – applies folk psychology to agents System 1 applies rules of goals and desires System 2 applies rules of belief and knowledge ToMM-2 Attitudinal agency

Baron-Cohen’s Model Eye Direction Detector (EDD) Intentionality Eye-like stimuli Self-Propelled Stimuli Eye Direction Detector (EDD) Intentionality Detector (ID) Dyadic representations (desire, goals) Dyadic Representations (sees) Shared Attention Mechanism (SAM) Requires two types of input stimuli: Eye-like stimuli Self-propelled (animate) stimuli Provides greater detail on the inner workings of Leslie’s ToMM-1 module Proposes that autism is an impairment of either SAM (subgroup A) or ToMM (subgroup B) Theory of Mind Mechanism (ToMM)

Pre-attentive filters Scassellati’s Model Animate Objects ToBY Trajectory Formation Face Finder Visual Attention EDD ID Pre-attentive filters f f SAM Visual Input Visual Input

Animacy Detection Animate Inanimate Animate

Elastic Collision Expert Acceleration Sign Change Expert Animacy Experts Straight Line Expert Minimize the sum of the deviations from the mean velocity Constant mass and the inertial system provides the gravity vector Energy Expert Elastic Collision Expert Look for transfer in velocities before and after collision Acceleration Sign Change Expert Look for multiple sign changes in the acceleration

Multi-Target Tracking

Trajectory Formation t t+1 t+2 t+3

Motion Correspondence (Reid 1979 • Cox and Hingorani 1996)

Mimicry

Faces and Eyes 300 msec 66 msec Locate target In wide field Foveate Apply Face Filter Software Zoom Feature Extraction 300 msec 66 msec

Mimicry

Research Issues Our Approaches K n o w i g h a t m e M p b d s C r c f l u z G x k I v S U y A D L P E - ( , ) Research Issues Our Approaches

Matthew Marjanović

“Trying to make a hard thing easier, By making it much harder.”

Virtual Musculature

Configurable Musculature

Virtual Metabolism

Hand

Giorgio Metta

Reaching

Cog’s Sensitive Side