chapter 7: hybrid deliberative/reactive paradigm

Introduction to AI Robotics (MIT Press), copyright Robin Murphy 2000 Chapter 7: Hybrid Deliberative/Reactive Paradigm 1

7Chapter 7:

Hybrid Deliberative/Reactive Paradigm

Part 1: Overview & Managerial ArchitecturesPart 2: State Hierarchy Architectures


7 Objectives

• Describe the hybrid paradigm in terms of 1) SPA and 2) sensing organization

• Given a list of responsibilities, be able to say whether it belongs in the deliberative layer or in the reactive layer

• List the five basic components of a Hybrid architecture: sequencer agent, resource manager, cartographer, mission planner, performance monitoring and problem solving agent.

• Be able to describe the difference between managerial, state hierarchy, and model-oriented styles of Hybrid architectures.

• Be able to describe the use of state to define behaviors and deliberative responsibilities in state hierarchy styles of Hybrid architectures


7 Motivating Example for Deliberation: USAR

• Worker places robot at entrance to unstable building, loads in the floor plan, contextual knowledge and tells robot to look for survivors efficiently and map out safe routes for workers to pass through

• contextual knowledge includes probability of where people are more likely to be

What can a reactive architecture do? What can’t it do?

Path planning, handling detours due to blockage, map making, learn from past rescues


7 Organization: Plan, Sense-Act

SENSE

PLAN

ACT


7 Sensing Organization

BEHAVIOR

BEHAVIOR

BEHAVIOR

SENSOR 1

SENSOR 2

ACTUATORS

WORLD MAP/KNOWLEDGE REP

SENSOR 3

virtual sensor

Deliberative functions*Can “eavesdrop”*Can have their ownSensors*Have output which Looks like a sensorOutput to a behavior(virtual sensor)


7 Deliberation v. Reactionas a function of TIME

• Past, Present, Future

• Reactive– exists in the PRESENT (will a bit of

duration)

• Deliberative– can reason about the PAST– can project into the FUTURE


7 Architectures:Key Questions

• How does the architecture distinguish between reaction and deliberation?

• How does it organize responsibilities in the deliberative portion?

• How does overall behavior emerge?


7 Architectures: Common Functionality

• Mission planner

• Cartographer

• Sequencer agent

• Behavioral manager

• Performance monitor/problem solving agent (fairly rare)


7 Architectures: 3 Styles

• Managerial (division of responsibilities looks like in business)– AuRA, SFX

• State Hierarchies (strictly by time scope)– 3T

• Model-Oriented (models serve as virtual sensors)– Saphira, TCA


7 Mgr Architecture 1:AuRA (Autonomous Robot Arch.)

Ron Arkin, Georgia Institute of Technology


7 AuRA Architectural Layout

reac

tive

del

iber

ativ

e


7 Architectures: Common Functionality

• Mission planner • Cartographer• Sequencer agent• Behavioral manager• Performance monitor/problem solving

agent


7 AuRA Architectural LayoutCartographer

Sequencer

MissionPlanner

Behavioral manager(mgr+schemas)

PerformanceMonitoring

Emergent behavior


7 HOW WOULD THIS DO USAR TASK?


7 Example USAR (overlay)• Cartographer accepts the map

• Navigator uses path planning algorithm to visit nodes in order of likelihood of survivors

• Pilot determines the list of behaviors, Motor Schema Manager instantiates them (MS & PS) and waits for termination

• Homeostatic might notice that robot is running out of power, so opportunistically picks up low probability room on way back to home


7 Mgr. Architecture 2:SFX (Sensor Fusion Effects)

• Focus on sensing

• Biomimetic organization

• deliberative layer consists of managerial agents

• reactive layer has tactical behaviors


7 SFX (Sensor Fusion Effects)


7 SFX (Sensor Fusion Effects)

Behaviors(using direct

perception, fusion)

SenseSenseSenseSenseMuscleMuscleMuscleActuators

Deliberative Layer Managers

SenseSenseSenseSensor

SenseSenseSenseReceptiveField

Choice of behaviors, resourceallocation, motivation, context

Focus of attention,recalibration

SensorWhiteboard

BehavioralWhiteboard

Del

iber

ativ

e L

ayer

Rea

cti v

e L

a yer

Parameters to behaviors,sensor failures, task progress

actions

SuperiorColliculus-likefunctions

CerebralCortex-likefunctions

Cartographer(model/map

making)

Recognitionperception


7

Cartographer

SFX Implementation

Sensor Mgr

Task Planner

Interface Mgr

BehaviorsBehaviorsBehaviors

SensorsSensors

SensorsSensors

Sensors

AcuatorsAcuators

Acuators

Effector Mgr

Lis

pC

++

Beh

avio

rsU

se, F

use

In

form

atio

nd

irec

ts s

ensi

ng

HOW WOULD THIS DO USAR?


7 Example USAR (overlay)

• Cartographer accepts the map

• Task Planner agent asks for path, requests behaviors, passes to managerial layer

• Sensing and Effector Mgrs negotiate allocation

• Behaviors run until terminate or encounter exception (either preset condition by mgrs or through monitoring)

• Mgrs can see “below” but not above--cannot relax constraint of Planner/Boss


7 Tactical Behaviors

sensors strategic behaviors tactical behaviors actuators

follow-path speed-controlcamera drive

motor

avoidsonar

steermotor

center-cameracamerapanmotor

inclino-meter

slope

clutter

obstacleshow much vehicle turns

direction to path safe direction

safe velocity

swivel camera

strategic velocity


7 Summary:Managerial Architectures

• How does the architecture distinguish between reaction and deliberation?

– Deliberation: global knowledge or world models, projection forward or backward in time

– Reaction: behaviors which have some past/persistence of perception and external state

• How does it organize responsibilities in the deliberative portion?– hierarchy of managerial responsibility, managers may be peer

software agents

• How does overall behavior emerge?– From interactions of a set of behaviors dynamically instantiated

and modified by the deliberative layer– assemblages of behaviors


7Chapter 7:

Hybrid Deliberative/Reactive Paradigm

Part 1: Overview & Managerial ArchitecturesPart 2: State Hierarchy & Model-Based Architectures


7 State Hierarchy Architectures

• How does the architecture distinguish between reaction and deliberation?– Deliberation: requires PAST or FUTURE knowledge– Reaction: behaviors are purely reflexive and have only local,

behavior specific; require only PRESENT

• How does it organize responsibilities in the deliberative portion?– By internal temporal state

• PRESENT (controller)• PAST (sequencer)• FUTURE (planner)

– By speed of execution

• How does overall behavior emerge?– From generation and monitoring of a sequence of behaviors– assemblages of behaviors called skills– subsumption


7 3T Architecture

• Used extensively at NASA

• Merging of subsumption variation (Gat, Bonasso), RAPs (Firby), and vision (Kortenkamp)

• Has 3 layers– reactive

– deliberative

– in-between (reactive planning)

• Arranges by time

• Arranges by execution rate– ex. vision in deliberation

Dave Kortenkamp,TRAC Labs (NASA JSC)


7


7cartographer

Mission planner

sequencer

Behaviormgr

Performancemonitor

Emergent behavior


7 3T architecture

• Skills have associated events, to verify that an action had correct effect.

• Skills operate only in the Present,

• components of the sequencer layer operate on state information about the Past, as well as present,planner layer works state information about the Past and Present to plan the Future,

• slow algorithms are in the Planner, fast algorithms go into the Skill Manager Layer, so vision algorithms were placed in the Planner despite their low-level sensing functions.


7 Model-Oriented Architectures• How does the architecture distinguish between reaction and

deliberation?– Deliberation: anything relating a behavior to a goal or objective– Reaction: behaviors are “small control units” operating in

present, but may use global knowledge as if it were a sensor (virtual sensor)

• How does it organize responsibilities in the deliberative portion?– Behavioral component– Model of the world and state of the robot– throwback to Hierarchical Paradigm with global world model

but virtual sensors– Deliberative functions

• How does overall behavior emerge?– From generation and monitoring of a sequence of behaviors– voting or fuzzy logic for combination


7 Model-Oriented Architectures• Top-down, symbolic flavor,• symbolic manipulation around a global world model,• world model supply perception with virtual sensors,• different than the hierarchical model:

- model restricted to labeling objects of interest like hallway, door, et, - perceptual processing is distributed and asynchronous - sensor errors and uncertaity can be filtered using sensor fusion over time to improve performance, - increase in processor speed and optimizing compilers solved the processing bottleneck


7Saphira Architecture• Developed at SRI by

Konolige, Myers, Saffioti

• Comes with Pioneer robots

• Behaviors produce fuzzy outputs, fuzzy logic combines them

• Has a global rep called a Local Perceptual Structure to filter noise

• Instead of RAPs, uses PRS-Lite


7 Saphira and LPS


7 PRS - Procedural Reasoning System

• Reactivity - postponement of the elaboration of plans until it is necessary,

• plans are but determined continuously in reaction to the current situation,

• plans in progress can be interrupted and abandoned at any time,

• plans represent the robot’s desired behavior,

• a symbolic plan always drives the system,


7 Sequencer agent,Mission Planning,Performance mon.

Cartographer

Behavior mgr

Emergent behavior


7 Saphira

• Three tenets of successful mobile robot architecture: - coordination of robot’s sensors, actuators and its goals over time ( not adressed by the reactive architecture), - coherence - ability of the robot to maintain global information about its world - essential for good behavioral performance and interaction with humans, - ability to community with other robots and humans

• deliberation activities distributed among software agents, some of them running on a local workstation,

• reactive components consists of behaviors,

• behaviors extract virtual sensor inputs from the Local Perceptual Space,

• behavioral output is fuzzy rules , which are fused together using fuzzy logic into a velocity and steer commands,


7 Task Control Architecture• Developed by Reid Simmons

• Used extensively by CMU Field Robotics Projects– NASA’s Nomad, Ambler,

Dante

• Closest to Hierarchical in philosophy, but strong reactive theme showing up in implementation


7 TCA

• closer to an operating system architecture,

• tasks instead of behaviors,

• uses dedicated sensing structures called evidence grids corresponding to a distributed global world model,

• Task Scheduling Layer (using Prodigy planner) determines the task flow, interacts with the user, determines the goals and order of execution,

• navigation handled by a Partially Observable Markov Decision Process (POMDP),

• Obstacle Avoidance Layer uses a curvature-velocity method to factor not only obstacles but how to respond with a smooth trajectory


7 TCA


7 Other Hybrid Architectures

• SSS - Servo Subsumption Symbolic - 3 layers, - world models are a convenience not a necessity, - symbolic (deliberative) layer switches behaviors on and off and provides parameters - symbolic layer - where to go next (strategic) - subsumption - where to go now (tactical)

• SOMASS Hybrid Assembly System - cognitive (deliberative)/subcognitive (reactive) - planner as ignorant as possible, hierarchical, - subcognitive components execute plan downloaded to the robot, - planning as configuration.


7 Other Hybrid Architectures

• Agent Architecture - plans are descriptions of intended behaviors, - two levels: physical (perception and action), cognitive (higher level reasoning), - functional boundary is blurred - reactive components can coexist within each level.

• Theo Agent - reacts when it can, plans when it must, - focusses on learning - how to become more reactive, more correct and more perceptive, - rule selection and execution, - when no rule available, new rule added by the planner


7 Evaluation of Hybrids

• Support of Modularity: high

• Niche targetability: high (ex. Lower levels of AuRA, SFX, 2 1/2 T is just reactive)

• Robustness: SFX and 3T explicitly monitor performance of the reactive behaviors and either replace or adapt the configuration as needed,

• better comply with the software engineering principles,

• global world models used only for sumbolic functions,

• frame problem reduced by the principle - “Think in closed world, act in open world”

• planners produce only partial plans asynchronously with the reactos


7 Hybrid Summary

• P,S-A, deliberation uses global world models, reactive uses behavior-specific or virtual sensors

• Architectures generally have modules for mission planner, sequencer, behavioral mgr, cartographer, and performance monitoring

• Deliberative component is often divided into sub-layers (sequencer/mission planner or managers/mission planner)

• Reactive component tends to use assemblages of behaviors

• Roots in ethology and a framework for exploring cognitive science.

chapter 7: hybrid deliberative/reactive paradigm

Documents

usar task

sfx sensor fusion effectsfocus

maptask planner agent

virtual sensor7deliberation

managerial layersensing

virtual sensorssaphira

list of behaviors

requests behaviors