Beyond Gazing, Pointing, and Reaching

A Survey of Developmental Robotics

Authors: Max Lungarella, Giorgio Metta

Overview Introduction

Research areas

Existing theories

Observations and future directions

Introduction What is developmental robotics?

Use robots to test models from developmental psychology and neuroscience

Applies insights from ontogenetic development

Why combine robotics with psychology? Novel methodologies New research tools

Aim of the article? Present state of developmental robotics Motivate use of robots as research tool

Research areas Criteria

Situatedness Addressing hypotheses raised by either

developmental psychology or neuroscience Order of identified research articles

Social Interaction Sensorimotor control Categorization Value systems Developmental plasticity Motor skill acquisition and morphological changes

Research area: social interaction What kind of social

interaction? Joint attention

Scassellati (1998, 2001) Nagai (2002)

Low-level imitation Demiris (1999)

Development of language Varshavskaya (2002)

Social regulation Dautenhahn and Billard

(1999)

Research area: sensorimotor control Crucial to interact with world Examples

Control of reaching Marjanovic (1996) Metta (1999)

Control of grasp Coehlo (2001)

Interaction with objects Metta and Fitzpatrick (2003)

Interaction with environment Berthouze (1996)

Research area: categorisation Categorisation in developmental robotics

How categories are formed By interaction with environment, searching for correlations

between sensors

Categorisation of objects Scheuer and Lambrinos (1996)

Sensorimotor related categorisation Berthouze and Kuniyoshi (1998)

Research area: value system Value systems in robotics

Internal mediator of environmental stimuli/events Used to guide exploration process

Value dependant learning Learning technique where value system alters the learning

by: Specifying mechanisms by which stimuli can modulate

learning Providing system with input that essentially is signal filtered

by agent’s value system Almassy (1998) – simulated neural model, value system

altered strength of connection from neurons of visual area to ones of motor area

Lungarella and Berthouze (2002) – value system used to explore parameter space

Research area: developmental plasticity Brain inspired

developing a brain is plastic (flexible) and the plasticity is experience dependent

Almassy (1998) Self generated movements crucial for emergence and

development of visual responses Foveal preference

Research area: morphological changes and motor skill acquisition Morphological changes

For example: body growth One of the most explicit characteristics of ongoing

developmental processes

Articles: Lungarella and Berthouze (2002)

How morphological changes influence acquisition of motor skills?

Does inherent adaptivity of motor development lead to behaviours not obtainable by simple value based regulation of neural parameters?

Comparative analysis between simultaneous and progressive use of available DOFs.

Simultaneous use of available DOFs reduces probability of physical entrainment.

Exisiting theories Developmental engineering:

Brooks and Stein (1991): development as way to construct intelligent robotic systems

Aim: “to show that adoption of framework of biological development is suitable for construction of artificial systems”.

Recognising long sequences of cause-effect relationships characterises learning in real context

Features of human-like intelligent systems (Brooks, 1998): Development Embodiment Social interaction Multisensory integrations

Key assumptions: Human intelligence not as general purpose as thought Intelligence does not require monolithic control system Intelligent agent does not require centrally stored model of real

world

Existing theories cont. Cognitive Developmental Robotics:

Asada (2001) Aim: “to avoid implementing robot’s control structure

according to designer’s understanding of robot’s physics, but to allow robot develop its own understanding”.

Robot no longer given externally designed structure

Autonomous Mental Development: Weng (2001) States that for robot to be truly mental developed means to

be non-task specific Aim: to develop robots that are non-task specific and able

to develop own task representation that could not be possibly embedded a priori by designer

Observations and future directions Majority of studies reviewed in paper belong

to either social interaction or sensorimotor control.

Researchers underline importance of developing robots with social and early motor competencies – very few try to achieve it.

Future direction: going beyond “gazing, pointing and reaching”