13 Robot Intelligence

7/29/2019 13 Robot Intelligence

1/40

Robots and

IntelligenceRichard M Crowder

November 2010


2/40

Some interesting questions.

How do ants forage

How do birds flock

How do we drive a car round a track without a driver

How can we use a bees vision system to land a helicopter


3/40

How do birds flock?

Why

Safety in numbers

Increased foraging efficiency.

How

Wants to move into the same direction as thesurrounding birds

Wants to be in the middle of the surrounding birdsWants to keep a certain distance from the surrounding

birds


4/40

Control of Robot systems

Initially robots were pre-programmed

Mechanical limit switches

Hard wired logic

Computers

Largely to follow as fixed task in manufacturing etc

Robotic research activity is now centred on behaviour based

systems

Wide range of control architectures


5/40

What is a control architecture

At its simplest level, it is the methodology of integrating anumber of software blocks to provide a robot with itsunique functionality

Approaches

NASRAM Deliberative

Subsumption Reactive

Hybrid


6/40

Robot control architecture

Key points:

Importance of coupling sensing and action

Avoidance of symbolic knowledge

Decomposition into meaningful units (behaviour etc) Between approach there are:

Differences in granularities

Behaviour specification

Encoding Coordination

Programming approach


7/40

Approaches to Robot Control

DELIBERATIVE REACTIVE

Symbolic Reflexive

Speed of response

Predictive Capabilities

Dependence on World Models

Representation dependantSlower responseHigh-level intelligenceVariable latency

Representation FreeReal-time responseLow-level intelligenceSimple computation


8/40

Key Issues

Grounding in Reality

Robots are grounded in the physical world problem withsimulation. Building robots crystallises ideas.

Concept of embodiment Ecological dynamics

Dynamic environment that changes is both space and time

Scalability

For example how do insects related to humans


9/40

Braitenburg Vehicles

++++

PhotophobePhotovore


10/40

Deliberative approach

Key features

Hierarchical in structure

Communication and Control is predictable

High level provide subgoals for lower levels


11/40

GlobalMemory

Sensory

processing

Value

judgement

World

mapping

Task

decomposition

Level 1

Global

Memory

Sensory

processing

Value

judgement

World

mapping

Task

decomposition

Level 2

Global

MemorySensory

processing

Value

judgement

World

mapping

Task

decomposition

Level 3

Global

Memory

Sensory

processing

Valuejudgement

World

mapping

Taskdecomposition

Level 4

Hierarchical Intelligent Control System

Global

MemorySensory

processing

Value

judgement

World

mapping

Task

decomposition

Level 5

Global

Memory

Sensory

processing

Value

judgement

World

mapping

Task

decomposition

Level 6

Servo

Mission

Planning


12/40

Standard model for teleoperated systems


13/40

Subsumption Architecture

Rodney Brookes mid 1980s

Considered that sense-plan-actis detrimental to theconstruction of real robots.

Building a world model and reasoning usingrepresentational knowledge is at best a impediment tospeed of response


14/40

Key points

Complex behaviour does not necessarily come fromcomplex control systems

Intelligence is in the eye of the observer

The world is its best model

Robots are cheap

Robustness is good (failing sensors)

Systems can be incrementally build


15/40

Foraging

Start

BEGIN

AcquireWander

Retrieve Halt

DETECT

GRAB

DONE

RELEASE

Wander: move through the world

Acquire: move towards an attractor when detectedRetrieve: Return to home


16/40

Approaches compared

Sense

Model

Plan

Act

Modify the world

Create maps

Discover new areas

Avoid collisions

Move around


17/40

Augmented finite state machine

BEHAVIORAL

MODULEINPUT

OUTPUT

Inhibitor

Reset

Suppressor


18/40

Three Layered Robot

WANDER

COLLIDE REVERSE

ResetLOST

FORWARD

GO

RUN

AWAYSensor

Clock

MOTORS

BRAKES

AVOID OBJECTS

EXPLORE

BACK OUT OF TIGHT CORNERS


19/40

Perception

The robots ability to interpret information about itsimmediate surroundings, and then react to the changes

The real world is hostile to robots

Things move and change without warning.

A Priori knowledge may be incorrect, inaccurate andobsolete


20/40

Need to consider the interaction between sensors andactuators

Perception without context of action is meaningless

How can the separation between reactive and deliberativecontrol system be resolved, to draw on their individualstrengths.


21/40

Original perception paradigm

Perception needs are

determined by the robots

motivation and behaviours


22/40

Revised perception paradigm

Moving

Object

Static

Object

Smooth

Surfaces

Vehicle

Time


23/40

Action Perception Cycle

Action Plans

BehaviourWorld

Modification

ModelsMemory

Reactive Shunt

This cycle is very much a human view of the interaction,

but the cycle can be reflected in robotics


24/40



Symbolic ReflexiveSpeed of response



Both approaches have weaknesses


25/40

Reactive Control

is a technique for tightly coupling perception and action, typically in thecontext of motor behaviour, to provide timely robotic responses in adynamic and unstructured world

Grounding in Reality

Robots are grounded in the physical world.

Concept of embodiment (an intelligent agent that interacts withthe environment through a physical body within that environment )

Ecological dynamics

Dynamic environment that changes is both space and time


26/40

but

A purely reactive systems makes a number of assumptions

The environment lacks temporal consistency and stability.

The robots sensing is adequate.

Difficult to localise a robot relative to the world model. Symbolic knowledge has little or no value.

Deliberative planning systems provide an entry point at which AI andsymbolic knowledge representation can enter a reactive system.

Evidence that hybrid systems are present in nature


27/40



Symbolic ReflexiveSpeed of response




28/40

Hybrid Control

Behaviour and perception can be configured to match thetask and environment.

A priori world knowledge, if stable, can be used to

reconfigure behaviours.

Dynamically acquired would knowledge can be used toavoid problems

.Hybrid control allows the reconfiguration of a reactivesystem based on world knowledge


29/40

Hybrid Architecture

ActuatorsSensors Control

Sequencing

Deliberative

Activation

Results

Status

Results

Atlantis Architecture


30/40

Features

Three layered approach

Asynchronous reactivity anddeliberation

Deliberation is viewed as advicenot decree

Failures provide opportunitiesfor restructuring

JPL Rocky 4


31/40

Sensors

Communicating with the world sensors can be viewed asa form on communication

The robot needs to pay attention to the key feature in the

image. How the attention and perception resources are used is a

function of motivation or intention.

Biological provides many examples:

Intraspecies kin recognition Prey detectors


32/40

SLAM

Simultaneous localization and mapping

Key problem in mobile robots

Build a map without a prioriknowledge and Keep track of where we the robot is

i.e

What does the world look like Where am I


33/40

Mobile Robot

Mobile platform fitted with odometry

Error limited to 2cm to 1m moved and 2o per 450 turned

Report position in X-Y coordinates

Laser scanner to locate landmarks etc.

Vision is over complex and causes a bottleneck in theprocess


34/40

Process overview

Laser

scanner

Data

association

Landmark

Extraction

odometry

change

EKF

odometry update

EKF

Re observation

EKF

New observation

EKF

Extended Kalman Filter

As the odometry is prone to error

it needs to be updated from the

environment


35/40

SLAM in practice (1)Landmark

Robot locates its positionagainst the landmarks


36/40


Robot moves and using odometryand determines the distance, but due to errors

the position is incorrect.

Actual

Estimated (O)


37/40


The landmark system relocates the robot,

as we believe this over the odometry, we

correct the odometry reading.

Actual

Estimated (O)

Estimated (L)


38/40

Landmarks

Landmarks should be easily re-observable.

Individual landmarks should be distinguishable from eachother.

Landmarks should be plentiful in the environment.

Landmarks should be stationary.

Examples: Room Corners


39/40

Problem with Landmarks

The system might not re-observe landmarks every timestep.

The system might observe something as being a landmark

but fail to ever see it again. The might wrongly associate a landmark to a previously

seen landmark.

Data Association solves this problem by storing ALL

landmarks seen, and using only those that have been seenN times


40/40

EKF

Extended Kalman Filter

Nonlinear version of the Kalman Filter

De facto standard for GPS and Navigation

Three main steps

Update the current state estimate using the odometrydata

Update the estimated state from re-observinglandmarks.

Add new landmarks to the current state.

Documents

13 Robot Intelligence