The Hierarchical Paradigm

2 The Hierarchical Paradigm• Describe the Hierarchical Paradigm in terms of the 3 robot primitives

and its organization of sensing

• Name and evaluate one representative Hierarchical architecture in terms of: support for modularity, niche targetability, ease of portability to other domains, robustness

• Solve a simple navigation problem using STRIPS (hint: work through Sec. 2.2.2)

• Understand precondition, closed world assumption, open world, frame problem

• List two advantages and disadvantages of the Hierarchical Paradigm

Introduction to AI Robotics (MIT Press)

Chapter 2: The Hierarchical Paradigm 1

Organization

-SPA

-global

Strips

-Shakey

Rep. Arch.

-evaluation

-NHC

-RCA

Summary

2 Hierarchical Paradigm…

• Top-down: – Plan, plan, plan

• Control-theoretic: – must measure error in order to control device

• Planning means:– dependence on world models



2 Organization



PLANSENSE ACT

World model:1. A priori rep2. Sensed info3. Cognitive

Organization

-SPA

-global

Strips

-Shakey

Rep. Arch.

-evaluation

-NHC

-RCA

Summary

2 Shakey• First AI robot

• Built by SRI (Stanford Research Institute) for DARPA 1967-9

• Used Strips as main algorithm for controlling what to do

Introduction to AI Robotics (MIT Press) Chapter 2: The Hierarchical Paradigm 4

Organization

-SPA

-global

Strips

-Shakey

Rep. Arch.

-evaluation

-NHC

-RCA

Summary

2 Strips: Means-ends analysis



INITIAL STATE: Tampa, Florida (0,0)

GOAL STATE: Stanford, California (1000,200)

Difference: 1020 miles

“Go to Stanford AI Lab”

Organization

-SPA

-global

Strips

-Shakey

Rep. Arch.

-evaluation

-NHC

-RCA

Summary

2 Difference Table



d>=200 miles FLY

100<d<200 TRAIN

d<=100 DRIVE

Distance(difference)

mode of transportation(OPERATOR)

d<1 WALK

mode=difference_table(INITIAL STATE, GOAL STATE, difference)

1. Look up what to do: FLY2. Not at SAIL, so repeat3. Look up what to do: DRIVE

Organization

-SPA

-global

Strips

-Shakey

Rep. Arch.

-evaluation

-NHC

-RCA

Summary

2 Preconditions



d>=200 miles FLY

100<d<200 TRAIN

d<=100 DRIVE (rental)

DRIVE (personal car)

difference OPERATOR

d<1 WALK

How do I know if I’m at the airport or at home?Now must keep up with the state of the world

at airport

at home

PRECONDITIONS

Organization-SPA

-globalStrips

-ShakeyRep. Arch.-evaluation

-NHC-RCA

Summary

2 Maintaining State of the World:Add and Delete Lists



d>=200 miles

FLY

100<d<200 TRAIN


at airport

DRIVE (personal)

at home

distance OPERATOR PRE-CONDITIONS

d<1 WALK

at city Yat airport

at city Yat train station

ADD-LIST

at city X

at city X

DELETE-LIST

Organization-SPA

-globalStrips


-NHC-RCA

Summary

2 Class Exercise

• Write down the world model, the operator applied, the change in world state, etc. to go from Tampa to Stanford

Introduction to AI Robotics (MIT Press) Chapter 2: The Hierarchical Paradigm 9

d<=200 miles

FLY

100<d<200

TRAIN


at airport

DRIVE (personal)

at home

distance OPERATOR PRE-CONDITIONS

d<1 WALK

at city Yat airport

at city Yat train station

ADD-LIST

at city X

at city X

DELETE-LIST

Organization-SPA

-globalStrips


-NHC-RCA

Summary

2 Strips Summary• Designer must set up

– World model representation– Difference table with operators, preconditions, add & delete

lists– Difference evaluator

• Strips assumes closed world– Closed world: world model contains everything needed for

robot (implication is that it doesn’t change)– Open world: world is dynamic and world model may not be

complete• Strips suffers from frame problem

– Frame problem: representation grows too large to reasonably operate over



Organization-SPA

-globalStrips


-NHC-RCA

Summary

2 Architecture• provides a principled way of organizing a control system.

However, in addition to providing structure, it imposes constraints on the way the control problem can be solved [Mataric]

• describes a set of architectural components and how they interact [Dean & Wellman]

• Types of architectures [Levis, George Mason University]– operational architecture: describes what the systems does, not how

it does it– systems architecture: describes how a system works in terms on

major subsystems– technical architecture: implementation details



Organization-SPA

-globalStrips


-NHC-RCA

Summary

2 Nested Hierarchical Controller(Meystel)



Organization-SPA-globalStrips-ShakeyRep. Arch.-evaluation-NHC-RCASummary

2 NHC Planner



Organization-SPA-globalStrips-ShakeyRep. Arch.-evaluation-NHC-RCASummary

2 RCS (Albus)

• the hierarchy

• how the hierarchy works for navigation

• how it is implemented– nodes and modules– planning time periods



2• hierarchy with node structure at each level

– have operator interface (in theory)

• nodes consist of– Sensory Processing– World Model– Behavior Generation– Value Judgment

• top-down, plan for a particular horizon– control theoretic



RCS (Albus)

2 Examples of RCS Apps



Organization

-SPA

-global

Strips

-Shakey

Rep. Arch.

-evaluation

-NHC

-RCA

Summary

2 Evaluating an Architecture• support for modularity: does it show good software

engineering principles?

• niche targetability: how well does it work for the intended application?

• ease of portability to other domains: how well would it work for other applications or other robots?

• robustness: where is the system vulnerable, and how does it try to reduce that vulnerability?



Organization-SPA

-globalStrips


-NHC-RCA

Summary

2 Evaluating NHC & RCS• support for modularity:

– decomposition by functionality• niche targetability:

– good, both have been used for apps like vehicle guidance, mining equipment

• ease of portability to other domains:

– unclear, not sure if code could be reused—lots of rewriting on previous apps

• robustness:

– RCS simulates plans in advance, but not sure what it would do with sensor or mechanical failures, etc.



Organization

-SPA

-global

Strips

-Shakey

Rep. Arch.

-evaluation

-NHC

-RCA

Summary

2 Advantages & Disadvantagesof Hierarchies

• Natural way to organize• Not basically rigid (it’s sort of flexible)• Not basically inefficient (sort of organized)

– not the same as centralized planning– priorities and goals are clear, therefore efficient



Advantage of Hierarchal ParadigmOrders the relationship between Sensing, Planning and Acting

2 Advantages & Disadvantagesof Hierarchies

• Planning bottleneck by updating cycle of the world model and the planning. - Sensing and planning algorithm were extremely slow. • Eliminating of stimulus response (urgent actions) - sensing and acting is disconnect. • Frame problem -> Strips• For Particular applications -> NHC• Never handled Uncertainty - semantics, sensor noise, actuator errors, action completion.



Disadvantages of Hierarchal Paradigm

2 Nodes are made from Modules



uff applies rule(transition rules)

sensors actions

u=uff+G(xd-x*)u is control action

x* is predicted world statexd is desired world state

uff is the feedforward control plan

transform into x*

G is feedback

xd is from “above”

2




sensors actions

if BALL, move toward centroidif NOT BALL, turn clockwise

(feedback determines how fast)

transform into x*G is feedbackxd is from “above”

Nodes are made from Modules

2




sensors actions

if BALL, move toward centroidif NOT BALL, turn clockwise

(feedback determines how fast)

transform into x*G is feedbackxd is from “above”

when to stop?how far is far enough?what about noise/fuzzy ball?…sensor noise, actuator error,rigid models

Nodes are made from Modules

2 Demo III XUV



http://museum.nist.gov/exhibits/timeline/item.cfm?itemId=38

Experimental Unmanned Vehicle in action at Ft. Indiantown Gap. Photo courtesy of the Army Research Labs. Nov. 2001

2 RCS-4 Levels



7 Battalion 2-24h

6 Platoon 5m-2h

5 Section 1-10m

4 Individual Vehicle 5-50s

3 Subsystem Level 200-500ms

2 Primitive Level 50-500ms

1 Servo Level 5-50ms

2 Each Level has a RCS Node



Sensory Processing, World Modeling,Behavior Generation, Value Judgment

Sensory Processing, World Modeling,Behavior Generation, Value Judgment

Engineering of Mind, Albus & Mystel, 2001

2 Implementation View:Nodes are Recursive



2 Demo III Control Hierarchy



PLANNERvehicle1 vehicle2

VEHICLE PLANNERcommunications plan AM plan RSTA plan

AM PLANNERDriver Plan Gaze plan

COMMS PLANNERmessage list

RSTA PLANNERgaze plan

DRIVER PLANNERVelocity Plan

GAZE PLANNERStereo Gaze Plan LADAR Gaze Plan

VELOCITY PLANNERF Wheels R Wheels F Steer R Steer

F Wheel R Wheel F Steer F Steer

Servo 50ms

Primitive 500ms

Subsystem 5s

Vehicle 1m

Section 10m

2 RCS XUV Example



Vehicle Level:AM Plan(A1…A10)

Primitive Level:Driver Plan(D1…D10)

2 t=0.5



Primitive Level:Driver Planextends to A2

2 t=1 s



Obstacle Detected

Primitive Level:Driver Plannew waypoints

Vehicle Level:detects too large avariation

2 t=1 s Vehicle Level Planner Opt 1



Vehicle Level:new AM Plan

Primitive Level:new Driver Plan

2 t=3



More obstacle is seen…fail upwards again

Vehicle Level:new AM Plan


2 t=4.5s



Vehicle Level:new AM Planskip A1, go to A2


2 t=6 s



Vehicle Level:new AM Planskip old A2


2 Exercise: Adapt to Rescue Robots?



PLANNERvehicle1 vehicle2

VEHICLE PLANNERcommunications plan AM plan RSTA plan

AM PLANNERDriver Plan Gaze plan

COMMS PLANNERmessage list

RSTA PLANNERgaze plan

DRIVER PLANNERVelocity Plan

GAZE PLANNERStereo Gaze Plan LADAR Gaze Plan

VELOCITY PLANNERF Wheels R Wheels F Steer R Steer

F Wheel R Wheel F Steer F Steer

Servo 50ms

Primitive 500ms

Subsystem 5s

Vehicle 1m

Section 10m

2 Hierarchical Review• Describe the hierarchical paradigm in terms of the three

robot primitives• Describe sensing in the hierarchical paradigm

• What is STRIPS?• What is the closed world assumption?• What are preconditions?• What is the frame problem?

• What are two representative architectures?• What is the NHC decomposition?



Organization

-SPA

-global

Strips

-Shakey

Rep. Arch.

-evaluation

-NHC

-RCA

Summary

The Hierarchical Paradigm

Documents

Transcript of The Hierarchical Paradigm