Sensorimotor Interaction in a Developing Robot

Post on 24-Feb-2016

49 views 0 download

Tags:

description

Sensorimotor Interaction in a Developing Robot. Metta , Sandini , Natale & Panerai. Introduction. Developmental principles based on biological systems. Time-variant machine learning. Focus on humanoid robots. Previous Work (to 2001). Some work in machine learning for robotics. - PowerPoint PPT Presentation

Transcript of Sensorimotor Interaction in a Developing Robot

Sensorimotor Interaction in a

Developing RobotMetta, Sandini, Natale & Panerai

Developmental principles based on biological systems.

Time-variant machine learning. Focus on humanoid robots.

Introduction

Some work in machine learning for robotics. Collect Data -> Train Machine -> Control

Robot Off-line training, tweaked by hand. Time-invariant

Previous Work (to 2001)

Physiology problem; explain how something in biology works.

A system is built by adapting from initial simplicity.

Non-adaptive systems often fail in the real world.

Real adaptation is hard to create and harder to control.

Principles of Development

Complex systems decomposed into small parts.

Parts are studied in isolation. Real world is not modular – newborns are

already integrated at birth. Not all ‘modules’ are functioning or at full

capabilities. All are matched and promote shift to more

complex behaviours.

Modularity vs. Integration

Example based learning is difficult to get right.

Basically function approximation. Too many parameters -> Overfitting Good approximation, bad generalisation. Too few -> Oversmoothing Bad approximation, no ‘grasp’ of problem

complexity.

Learning by Example (or not)

Control the complexity and structure of the learner.

Different from learning which controls parameters of the structure.

Better to start with a simpler system. Training data has a cost – exploration. Failure is not an option!

Development

Initial reflex-like starting conditions bootstrap the system.

Gather data through action, but without incurring penalties.

Quality of data linked to how the system acts.

Perception is derived from action. Not just sensory processing.

Development cont.

Mirror Neurons Found in the frontal cortex. Activated when an action is performed and

seen. Canonical Neurons Responsive to actions like grasping. Also respond to seeing a graspable object.

Neurons

Assume a limited set of skills and motor control abilities.

Build new abilities on top of old ones. Learn -> Act -> Perceive (randomly)

Actions Build on Actions

Actions must have consequences. Relate movements to sensorial

consequences. Eye and head tracking first develops

synchronisation, then tunes the amplitude of the movements.

A Perception of Causality

Objects are classified by what you can do to them.

Learn affordances by action. Measure outcome at sensory level. Grasping is learnt because possession is

‘good’.

Affordances

12 degrees of freedom.

Cameras, microphones, inertial sensors.

Orienting and reaching toward objects based on vision or audition.

Babybot

Reflex grasping as the robot learns to control gaze direction.

Gradual mapping between sound, vision and grasping.

Performs better with initially restricted vision that develops.

Babybot’s Development