The Reactive Paradigm

Introduction to AI Robotics (MIT Press) Chapter 4: The Reactive Paradigm 1

4 The Reactive Paradigm

• Describe the Reactive Paradigm in terms of the 3 robot primitives and its organization of sensing

• List the characteristics of a reactive robotic system, and discuss the connotations of surrounding the reactive paradigm

• Describe the two dominant methods for combining behaviors in a reactive architecture: subsumption and potential field summation

• Be able to program a behavior using pfields

• Be able to construct a new potential field from primitive pfields and sum pfields to generate an emergent behavior

ReviewOrganization-SA-beh. specificSubsumption-Philosophy-Level 0-Level 1-Level 2Summary


4 Review: Lessons from Biology

• Programs should decompose complex actions into behaviors. Complexity emerges from concurrent behaviors acting independently

• Agents should rely on straightforward activation mechanisms such as IRM

• Perception filters sensing and considers only what is relevant to the task (action-oriented perception)

• Behaviors are independent but the output may be used in many ways including: combined with others to produce a resultant output or to inhibit others



4 Hierarchical Organization is“Horizontal”



4 More Biological is “Vertical”


Higher level behaviors reuse or

inhibit more primitive behaviors


4 Sensing is Behavior-Specific or Local

Behaviors can “share” perception without knowing itThis is behavioral sensor fusion



4 Reactive Robots

• Most apps are programmed with this paradigm

• Biologically based:– Behaviors (independent processes), released by perceptual or internal

events (state)

– No world models or long term memory

– Highly modular, generic

– Overall behavior emerges

SENSE ACT

RELEASERbehavior

OverviewHistoryReactive USARSummary


4 Example 1: Robomow

• Behaviors?

• Random

• Avoid– Avoid(bump=obstacle)

– Avoid(wire=boundary)

• Stop– Stop(tilt=ON)

• All active www.friendlymachines.com



4 Example 2: My Real Baby

• Behaviors?

• Touch-> Awake

• Upside down & Awake-> Cry

• Awake & Hungry -> Cry

• Awake & Lonely -> Cry

• Note can get crying from multiple behaviors

• Note internal state (countdown timer on Lonely)

www.irobot.com



4 Reactive Behaviors• Connotations

– Execute rapidly • Can be implemented in hardware

– Have no memory • Characteristics of reactive architectures

– Robots are situated agents operating in an ecological niche– Behaviors are basic building blocks for robotic actions, and overall

behavior of robot emerges from their interaction• Independent, concurrent• Schema of a behavior may have a coordinated control program, but there

is no external controller of all behaviors for a task• Architecture may use combination, suppression, or cancellation for

interaction– Only local, behavior-specific sensing in permitted

• No world model, representation from sensing is ego-centric– Modular

• Tasks are broken down into behaviors; behaviors can be tested independently; new behaviors can be built from more primitive ones

– Based on animal models of behavior


4 Reactive

• Historically, there are two main styles of creating a reactive system– Subsumption architecture

• Layers of behavioral competence

• How to control relationships

– Potential fields• Concurrent behaviors

• How to navigate

• They are equivalent in power

• In practice, see a mixture of both layers and concurrency


4 Subsumption:Rodney Brooks

From http://www.spe.sony.com/classics/fastcheap/index.html



4 Subsumption Philosophy• Modules should be grouped into

layers of competence

• Modules in a higher lever can override or subsume behaviors in the next lower level

– Suppression: substitute input going to a module

– Inhibit: turn off output from a module

• No internal state in the sense of a local, persistent representation similar to a world model.

• Architecture should be taskable: accomplished by a higher level turning on/off lower layers



4 Level 0: runaway

HALT

COLLIDE

PS MS

RUN AWAYPS MS

runaway 0

wander 1 follow-corridor 2


sensors motor actionsbehaviors

when obstacle comes near turn around, run away;when collision imminent, stop, turn around, run away

direction

magnitude


4 Example Perception: Polar Plot

• Plot is ego-centric

• Plot is distributed (available to whatever wants to use it)

• Although it is a representation in the sense of being a data structure, there is no memory (contains latest information) and no reasoning (2-3 means a “wall”)


if sensing is ego-centric, canoften eliminate need for memory, representation


4 Level 1: Wander

runaway 0



encoders

AVOID

PS

MS

WANDER

PS MS

Note sharing ofPerception, fusion

Avoid suppresses

(replaces)o

utput from

runawayWhat wouldInhibition do?


4 Class Exercise

runaway 0

wander 1 move2light 2

LIGHTPHOTO-

TROPHISM

S



4 Level 2: Follow-Corridors

runaway 0


STAY-IN-MIDDLE

PS MSReviewOrganization-SA-beh. specificSubsumption-Philosophy-Level 0-Level 1-Level 2Summary

Computed from shaft encoders

Intended course


4 Class Exercise

• Design the roomba with subsumption


4 Subsumption Review• What is the Reactive Paradigm in terms of primitives?

– Sense, act• What is the Reactive Paradigm in terms of sensing?

– Local to each behavior• Does the Reactive Paradigm solve the Open World problem?

– Open world is non-monotonic; need truth maintenance mechanism– Reactive paradigm has no memory, no truth maintenance

• How does the Reactive Paradigm eliminate the frame problem?– No world model, so no frame problem

• What is the difference between a behavior and a level of competence?– Not schema theoretic; level of competence groups schema-like

modules into abstract behaviors• What is the difference between suppression and inhibition in

subsumption?– Suppression acts like a gate; inhibition like on/off switch



4 Potential Fields:Ron Arkin

From http://www.cc.gatech.edu/aimosaic/faculty/arkin

From http://www.cc.gatech.edu/aimosaic/robot-lab/MRLhome.html


4 Potential Fields Philosophy• The motor schema component of a behavior can be

expressed with a potential fields methodology– A potential field can be a “primitive” or constructed from

primitives which are summed together– The output of behaviors are combined using vector summation

• From each behavior, the robot “feels” a vector or force– Magnitude = force, strength of stimulus, or velocity– Direction

• But we visualize the “force” as a field, where every point in space represents the vector that it would feel if it were at that point


4 Example: Run Away via Repulsion


4 5 Primitive Potential Fields

uniform

perpendicular

tangential


4 Draw These Now!Common fields in behaviors

• Uniform– Move in a particular direction, corridor following

• Repulsion– Runaway (obstacle avoidance)

• Attraction– Move to goal

• Perpendicular– Corridor following

• Tangential– Move through door, docking (in combination with other fields)

• random– do you think this is a potential field? what would it look like?


4 Class Exercise

• Name the field you’d use for – Moving towards a light

– Avoiding obstacles

Attractive

Repulsive


4 Magnitude profiles

• Constant magnitude

• Linear drop off

• Exponential drop off

• P. 127


4 Programming a potential field

Example: repulsive field with linear dropoff, only one sensor

• Vdirection = -180 degrees• Vmagnitude = (D-d)/D if d <= D where D is range of potential field

0 otherwise

(magnitude controls velocity)

Implementation in C on p. 129-130

With more than one sensor – fig. =4.18, p. 134 - 136


4 Problems

• Impact of update rates– if time between updates is too long

• Path can be jerky

• Can overshoot

• Robots can’t change velocity and direction immediately

• Fields may sum to 0


4 Combining Fields forEmergent Behavior

obstacleobstacle

goal

If robot were dropped anywhere on this grid,it would want to move to goal and avoid obstacle:

Behavior 1: MOVE2GOALBehavior 2: RUNAWAY

The output of each independent behavior is a vector,the 2 vectors is summed to produce emergent behavior

obstacle

goal


4

Note: In this example, repulsive field only extends for 2 meters;the robot runs away only if obstacle within2 meters

Note: in this example, robot can sense thegoal from 10 meters away

Fields and Their Combination


4 Path Taken

• If robot started at this location, it would take the following path

• It would only “feel”the vector for the location, then move accordingly, “feel” the next vector, move, etc.

• Pfield visualization allows us to see the vectors at all points, but robot never computes the “field of vectors” just the local vector

Robot only feels vectors for this

point when it (if) reaches that point


4


4 Discussion

• Could you represent the Arctic Tern feeding behavior with potential fields?– what happens with multiple red dots?

– can you inhibit with potential fields?


4 Example: follow-corridor or follow-sidewalk

Perpendicular Uniform

Combined

Note use of Magnitude profiles:Perpendicular decreases


4 Class Exercise:Draw Fields for Wall-Following(assume that robot stands still if no wall)

Just half of a follow-corridor, but…


4 But how does the robot see a wall without reasoning or intermediate

representations?

• Perceptual schema “connects the dots”, returns relative orientation

PS:Find-wall

MS: Perp.

MS: UniformS

Sonars

orientation


4 OK, But why isn’t that a representation of a wall?

• It’s not really reasoning that it’s a wall, rather it is reacting to the stimulus which happens to be smoothed (common in neighboring neurons)


4 Level 0: Runaway

Note: multiple instances ofa behavior vs. 1:Could just have 1 Instance of RUN AWAY,Which picks nearest reading;Doesn’t matter, but thisAllows addition of anotherSonar without changing theRUN AWAY behavior


4 Level 1: Wander

Wander isUniform, but

Changes directionaperiodically


4 Level 2: Follow Corridor

Follow-corridor

Should weLeaveRun AwayIn? Do weNeed it?


4 Pfields

• Advantages– Easy to visualize

– Easy to build up software libraries

– Fields can be parameterized

– Combination mechanism is fixed, tweaked with gains

• Disadvantages– Local minima problem (sum to magnitude=0)

– Jerky motion


4 Example: Docking Behavior

•Arkin and Murphy, 1990, Questa, Grossmann, Sandini, 1995, Tse and Luo, 1998, Vandorpe, Xu, Van Brussel, 1995. Roth, Schilling, 1998, Santos-Victor, Sandini, 1997

Orientation, ratio of pixel counts tangent vectorTotal count attraction vector

Selective attraction field ,

width of +-45 degrees


4 Docking Behavior Video


4 Class Discussion:When Does a Field End?

• Imagine the case of a “SodaPup” robot (MIT)– task: find and pick up a Coca-Cola can

– environment: red cans are only red object in world


4 Pfield advantages

• Easy to visualize

• Easy to combine

• Can be parameterized (different ranges, drop off, etc.)


4 disadvantages

• Local minima– Solutions:

• Add noise

• Navigation templates

• Express potential fields as harmonic functions – no local minima of 0, but computationally expensive


4 Pfields Summary• Reactive Paradigm: SA, sensing is local

– Solves the Open World problem by emulating biology

– Eliminates the frame problem by not using any global or persistent representation

– Perception is direct, ego-centric, and distributed

• Two architectural styles are: subsumption and pfields

• Behaviors in pfield methodologies are a tight coupling of sensing to acting; modules are mapped to schemas conceptually

• Potential fields and subsumption are logically equivalent but different implementations

• Pfield problems include– local minima (ways around this)

– jerky motion

– bit of an art

The Reactive Paradigm

Documents

Transcript of The Reactive Paradigm