Mobile & service robotics 03 - PoliTO

Mobile & Service RoboticsMobile & Service Robotics

Supervision and ControlSupervision and ControlSupervision and ControlSupervision and ControlBasilio Bona ROBOTICA 03CFIOR 1

Traditional approach

Traditional artificial intelligence considers robot “brains” as sequential processing units

Main ideasRepresentation, planning, reasoningModel building (geometrical maps of the environment)g (g p )Functional decomposition; hierarchical systemsSymbolic manipulation

Basilio Bona 2ROBOTICA 03CFIOR

Symbolic manipulation

Supervision and Control

L li ti & Path planning

PositionGlobal map

a priori knowledge Task/mission commands

Localization &Map Building

Path planning& Reasoning

Data Data n olData

treatmentData

treatment

datarcep

tion

on c

ontr

o

commands

iteration

Sensors Actuators

data

Per

Motio

Environment


Environment

Supervision and Control

P itiLocalizationMap Building

Path planningReasoning

PositionGlobal map

Local mapWorld model

Path

Perception Motioncontrolp control

Environment


Control Strategies

Structure of the control loopWorld changes dynamically

d l f h ld d ff l d fA compact model of the world is very difficult to defineThere are many sources of uncertainty, both in the world and in the robotrobot

Two possible approachesClassic AI – deliberativedeliberative model

Approximate world modeling (model‐based method)Based on a set of functions Vertical decomposition Top‐down approach

Modern AI reactivereactive modelModern AI – reactivereactive modelNo world model is given: behavior‐based controlHorizontal decompositionHorizontal decompositionBottom‐up approach


Control Characteristics

Sense Plan Act Subsumption/Reactive modelSense – Plan – Act

This architecture may prevent a fast and timely response

Subsumption/Reactive model

y p

sense

ch

Function 1

ch

use model

appro

ac Function 2

Function 3lappro

a

plan

Ver

tica

l a Function 3

Function 4orizo

nta

act

V

Function 5

Ho


Model‐Based Approach

Complete modeling of the worldEach block is a computed functionpVertical decomposition and nested‐embodiment of functions

An example

Perceptionsensors

Localization - Map building

Cognitive planning

Motion control actuators

Basilio Bona ROBOTICA 03CFIOR 7

Model‐Based Approach

Planner GOAL RECOGNITION

Navigator

GLOBAL PATH PLANNING

SUB-GOAL FORMULATION

Pilot

LOCAL PATH PLANNING

TARGET GENERATOR

Path monitor

DYNAMIC PATH PLANNING

TARGET LOCATION

Controller

PATH CORRECTION/OBSTACLE AVOIDANCE

COMMANDS

Low level controlSENSORS

ACTUATORS


ACTUATORS

Behavior‐Based Approach



Reactive systemsReflexive behavior Perception‐actionSubsumptionSubsumption

ROBOTROBOT

Perception 1Perception 2 Action 1Action 2

WORLDWORLD



No model is necessaryHorizontal decompositionpCoordination + Priority = FusionBiomimesis = observe and copy animal behaviorBiomimesis observe and copy animal behaviorSubsumptionEmbodimentEmbodiment

Avoid obstacles

Discover new areas

Detect goal positionsensors actuators∑Communicate data

Recharge

∑


Follow right/left wall

Coordination / fusion


Rodney Brooks is the inventor of this approach: Complex behavior needs not necessarily be the product of a complex control systemThe world is its best modelSi li i i iSimplicity is a virtueIntelligence is in the eye of the observerRobots should be cheapRobots should be cheapRobustness in the presence of noisy or failing sensors is a design goalPlanning is just a way of avoiding figuring out what to do nextPlanning is just a way of avoiding figuring out what to do nextAll onboard computation is importantSystems should be built incrementallySystems should be built incrementallyNo representation. No calibration. No complex computers. No high band communication








Subsumption architecture

The subsumption architecture was originally proposed by Brooks [1986]The subsumption architecture copies the synergy between sensation and actuation in lower animals such as insectsBrooks argues that instead of building complex agents in simple worlds, we should follow the evolutionary path and start building simple agents in the real, complex and unpredictable worldFrom this argument, a number of key features of subsumption result









1. No explicit knowledge representation is used. Brooks often refers to this as “The world is its own best model”

2 Behavior is distributed rather than centralized2. Behavior is distributed rather than centralized3. Response to stimuli is reflexive – the perception‐action sequence is

not modulated by cognitive deliberationy g4. The agents are organized in a bottom‐up fashion. Thus, complex

behaviors are fashioned from the combination of simpler, d l iunderlying ones

5. Individual agents are inexpensive, allowing a domain to be populated by many simple agents rather than a few complex onespopulated by many simple agents rather than a few complex ones. These simple agents individually consume little resources (such as power) and are expendable, making the investment in each agent i i lminimal

6. Several extensions have been proposed to pure reactive subsumption systems These extensions are known as behavior‐subsumption systems. These extensions are known as behaviorbased architectures


Control Strategies

DELIBERATIVEModel-based

REACTIVEBehavior-based

Speed of response

Purely symbolic Reflexive

Predictive capabilitiesPredictive capabilities

Depends on accurate world modelsDepends on accurate world models

• Depends on the world • Representation-freerepresentation

• Slow response• High level intelligence (cognition)• Variable latency

• Real-time response• Low level intelligence (stimulus-

response)• Fast and easy computation


Embodiment

To embody (verb) = to manifest or personify in concrete form; to incarnate; to incorporate, to unite into one body Embodiment is the way in which human (or any other animal)Embodiment is the way in which human (or any other animal) psychology arises from the brain & body physiologyEmbodiment theory was introduced into AI by Rodney Brooks in the y y y‘80s. Brooks have claimed that all autonomous agents need to be both embodied and situatedTh th t t th t i t lli t b h i f thThe theory states that intelligent behavior emerges from the interplay between brain, body and world. Brain, body and world are equally important factors in the explanation of how particular q y p p pintelligent behaviors originate in practiceBrooks showed that robots could be more effective if they “thought” ( l d d) littl ibl(planned or processed) as little as possible The robot's intelligence is organized only for handling the minimal amount of information necessary to make its behavior beamount of information necessary to make its behavior be appropriate and/or as desired by its creator


Embodiment

Rolf Pfeifer (AILab Zurich) says that there are essentially two directions in artificial intelligence:

one focused on developing useful algorithms or robots; and another direction that focuses on understanding intelligence, biological or artificial

In order to make progress on the second direction, an embodied perspective is mandatory


Situated robot

A situated robot is one which does not deal with abstract representations of the world (which may be simulated or real), but rather reacts directly to its environment as seen through its sensorsAn alternative to having a situated robot would be one which builds up a representation of its world and then makes plans b d hbased on this representationBecause of the limitations of our present technology, these two approaches often seem contradictoryAt present, each approach can be appropriate for different applications


Situated robot

The situated approach is good for dealing with problems whereThe situated approach is good for dealing with problems where planning ahead is unnecessary or takes too much time

However the representation approach is needed for solving moreHowever, the representation approach is needed for solving more complicated problems where it is necessary to reason about the state of the world

For dealing with complicated tasks in the real world, it will probably be necessary to fuse the two approaches

Reasoning can be used to build up higher level plans and solve high level problems

Lower level functions may use a more situated approach for carrying out plans and dealing with problems which need immediate attentionattention

The structure and relations that originates from the interaction of simple controllers and complex environment is called emergentsimple controllers and complex environment is called emergent behavior


Robotics and AI

Main areas of Artificial Intelligence applied to robotics1. Knowledge representation2. Understanding natural languages3. Learning4. Planning and problem solving5. Inference6. Search7. Vision


Knowledge representation

Define, build and memorize the physical and virtual structures used by the robot to represent

the worldthe desired tasksitself

Example: a robot is looking after a human being under the p g gwreckage of a fallen building: how it is represented?

Structural model:Head (oval) + trunk (cylindrical) + arms (cylindrical)Bilateral symmetry

Physical model (thermal image, …)What happens if the body is only partially visible?


Understanding natural languages

Natural language is one of the most simple and human ways to interact

But …To understand the words does not mean to understand the meaning

Grammatical structure vs Semantic structureGrammatical structure vs Semantic structure

ExampleWe ga e the monke s t o bananas beca se the e e h ngWe gave the monkeys two bananas because they were hungryWe gave the monkeys two bananas because they were over-ripe

They have the same grammatical structure, but a very different semantic structure

To understand the sense we must know both the monkeys and the bananas


Necessity to develop ontologies

Ontology

An ontology is a formal representation of knowledge as a set of concepts within a domain, and the relationships between those

I i d b h i i i hi h d iconcepts. It is used to reason about the entities within that domain, and may be used to describe the domain

An ontology is a “formal, explicit specification of a shared conceptualization.”An ontology provides a shared vocabulary, which can be used to model a domain — that is the type of objects and/orcan be used to model a domain — that is, the type of objects and/or concepts that exist, and their properties and relations

Ontologies are the structural frameworks for organizing informationOntologies are the structural frameworks for organizing information and are used in artificial intelligence, etc., as a form of knowledge representation about the world or some part of itp p


Learning

Learning is the capacity to memorize actions and behaviors and to repeat them to adapt to the implicit or explicit objectivesI b d l i i th bilit t d t d i lifIn a broad sense, learning is the ability to adapt during life We know that most living organisms with a nervous system display some type of adaptation during lifesome type of adaptation during lifeThe ability to adapt quickly is crucial for autonomous robots that operate in dynamic and partially unpredictable environments, but the learning systems developed so far have so many constraints that are hardly applicable to robots that interact with an environment without human interventionwithout human interventionLearning requires

A structure able to store and retrieve dataOne or more explicit objectivesAn adaptation mechanism (reward + punishment) An explicit or implicit teacher


Planning and problem solving

Intelligence is associated to the ability to plan actions toward the the given task fulfillment, and to solve problems arising when plans fail

Go thereGo there


Planning and problem solving


SLAM

SLAM Simultaneous Localization and Mapping


Inference

Inference is a procedure that allows to generate an answer alsoInference is a procedure that allows to generate an answer also when the available data or information are incomplete

Inference is based on statistical models (bayesian networks) orInference is based on statistical models (bayesian networks) or semantic models


Search

Search does not necessarily mean a true search of objects in space, but defines the ability to examine a knowledge

d b ( h ) f d h drepresentation data‐base (search space) to find the required answer

d l h h b f dConsider a computer playing chess: the best move is found looking for a solution in the search space of all possible moves, starting from the present chessboard statestarting from the present chessboard state


Vision

Vision is the most important sense in human beingsVision is the most important sense in human beings

Psychological studies have demonstrated that the ability to solve problems is due to our brain capacity to visualize thesolve problems is due to our brain capacity to visualize the effects of each action


Books

R.C. ArkinBehavior Based Robotics

G. Dudek, M. JenkinComputational Principles ofBehavior-Based Robotics

MIT Press, 1998

Computational Principles ofMobile RoboticsCambridge U.P., 2000

R.R. MurphyIntroduction to AI RoboticsMIT P 2000

R. Siegwart, I.R. NourbakhshAutonomous Mobile RobotsMIT Press, 2004


MIT Press, 2000

Books

Rolf Pfeifer, Josh C. BongardH th B d Sh th W WHow the Body Shapes the Way WeThink A New View of IntelligenceForeword by Rodney BrooksMIT Press 2006

S. Thrun, W. Burgard, D. Fox

MIT Press, 2006

S. Thrun, W. Burgard, D. Fox Probabilistic RoboticsMIT Press, 2005

Autori VariPrinciples of Robot Motion


Principles of Robot MotionMIT Press, 2005

Mobile & service robotics 03 - PoliTO

Documents