Sistemi per il Governo dei Robot

(1)

Sistemi per il Governo dei Robot

Silvia Rossi - Lezione 10

(2)

PRINCIPI E PROBLEMI NEL TRASFERIRE I SUGGERIMENTI AI ROBOTS

Riassumere dei principi generali di intelligenza naturale può essere utile per programmare robot:

I programmi dovrebbero decomporre azioni complesse in behaviour indipendenti che strettamente accoppiano percezione e azione.

I behaviour sono intrinsecamente paralleli e distribuiti.

Per semplificare il controllo e la coordinazione dei

behaviour, un agente dovrebbe contare su un corretto meccanismo di attivazione booleana (e.g. IRM).

Per semplificare la percezione questa dovrebbe filtrare i percetti e dovrebbe considerare quello che è importante per il behaviour (percezione action-oriented).

martedì 13 novembre 12

(3)

La percezione diretta (l'affordances) riduce la complessità computazionale del percepire e

permette alle azioni di esplicarsi senza memoria, inferenze, o interpretazioni.

I behaviour sono indipendenti, ma l'output di ciascuno

1) può essere combinato con un altro per produrre un output risultante, (cooperazione)

2) può servire ad inibire un altro (competizione).

(4)

ANIMAL BEHAVIOR

ROBOT EXAMPLES

(5)

Implicit Chaining

Reflexive behaviors are independent, not explicitely chained together.

Overall behavior emerges

(6)

Case Study: Lobsters and Odor Gated Rheotaxis

Taxis: orienting to a stimulus

Rheotaxis orienting to a flow

Lobsters

Live in turbulent water (energy intensive)

are very efficient at finding food via odor in water

(7)

Chemical Plumes & Chemotaxis

Simple “I’m in the plume, I’m out the plume” or

“move in direction of highest increase in

(8)

Hungry Lobster

If detect odor, moves in the direction of mean flow of the water

If loses odor, casts either to the right or the left randomly

If casting in one direction, after a while, will cast in the other direction

If doesn’t regain the plume, then quits after a while

(9)

Affordances

If detect odor, moves in the direction of mean flow of the water

If loses odor, casts either to the right or the left randomly

If casting in one direction, after a while, will cast in the other direction

If doesn’t regain the plume, then quits after a while

(10)

Absence of an Affordance

If detect odor, moves in the direction of mean flow of the water

If loses odor, casts either to the right or the left randomly

If casting in one direction, after a while, will cast in the other direction

If doesn’t regain the plume, then quits after a while

(11)

Absence of an Affordance

If detect odor, moves in the direction of mean flow of the water

If loses odor, casts either to the right or the left randomly -> FIXED-ACTION PATTERN

If casting in one direction, after a while, will cast in the other direction

If doesn’t regain the plume, then quits after a while

(12)

Doesn’t Have Memory But...

If detect odor, moves in the direction of mean flow of the water

If loses odor, casts either to the right or the left randomly

If casting in one direction, after a while, will cast in the other direction

If doesn’t regain the plume, then quits after a while

(13)

Doesn’t Have Memory But...

If detect odor, moves in the direction of mean flow of the water

If loses odor, casts either to the right or the left randomly

If casting in one direction, after a while, will cast in the other direction -> INTERNAL STATE

If doesn’t regain the plume, then quits after a

while -> INTERNAL STATE

(14)

Could Implement as FSA

HUNGRY

FEEDING

MOVE

UPSTREAM

CAST

odor detected

time passed

food found

odor regained

odor lost

(15)

(16)

Implicit Chaining v. Explicit

Explicit

Rapresentation on states

Doesn’t capture hungry or strength of smell may influence how long you cast or the intensity

(stimulus intensity)

(17)

Strenght of Stimulus

Consider attraction to red

Bigger area of red, may be sharper turn towards it (eg., higher gain)

Strength is either continous or discrete Linear (1/d)

Inverse square (1/d^2)

(18)

Cockroach Locomotion

• Case Western Biologically Inspired Robotics Laboratory (Roger Quinn)

• Studied mechanisms of locomotor behavior in American cockroach

• Developed a neural model faithful to biology:

– Uses cell membrane properties – Synaptic currents

– Generates outputs in terms of neuron’s firing frequency

• In simulation studies, achieved spontaneous generation of gaits observed in natural insect

• Behaviors included:

– Wandering

– Edge following

– Appetitive orientation and attraction to food

– Fixed-action pattern representing food consumption

ROBOT III

(19)

Reflexive Cockroach hiding

light goes on, the cockroach turns and runs when it gets to a wall, it follows it

when it finds a hiding place, goes in and faces outward

waits, then comes out

even if the lights are turned back off earlier

(20)

Fixed Pattern Action

light goes on, the cockroach turns and runs when it gets to a wall, it follows it

when it finds a hiding place, goes in and faces outward

waits, then comes out

even if the lights are turned back off earlier

(21)

Exhibits Taxis

light goes on, the cockroach turns and runs when it gets to a wall, it follows it

when it finds a hiding place, goes in and faces outward

waits, then comes out

even if the lights are turned back off earlier

(22)

Exercise

How many behaviors are there?

light goes on, the cockroach turns and runs when it gets to a wall, it follows it on right

when it finds a hiding place, goes in and faces outward

waits until not scared, then comes out

even if the lights are turned back off earlier

(23)

Break into Behaviors

Flee

light goes on, the cockroach turns and runs

Follow-wall

when it gets to a wall, it follows it on right

Hide

when it finds a hiding place, goes in and faces outward waits until not scared, then comes out

(24)

Releasers

Flee

light goes on, the cockroach turns and runs

Follow-wall

when it gets to a wall, it follows it on right

Hide

when it finds a hiding place, goes in and faces outward waits until not scared, then comes out

(25)

Need Internal State

Flee

light goes on, the cockroach is scared turns and runs

Follow-wall

when it gets to a wall and is still scared, it follows it on right

Hide

when it is scared and finds a hiding place, goes in and faces outward waits until not scared, then comes out

(26)

What Regulates Internal State?

Homeostasis (hunger) Affect (emotion)

Time (wears out or give up) ...

(27)

Need Internal State

Flee

light goes on, the cockroach is scared, turns and runs

Follow-wall

Hide

(28)

Sensig to Enable Behavior

Flee

light goes on, the cockroach is scared, turns and runs

Follow-wall

Hide

(29)

Emergent Linkage

What happens when lights are off, not scared, but at wall?

You don’t need to explicitly link behaviors!

(30)

Another Simplified Model of Cockroach Behavior

SENSORS BEHAVIORS MOUTH TACTILE

MOUTH CHEMICAL ANTENNA CHEMICAL

ANTENNA TACTILE

INGESTING

FINDING FOOD

EDGE FOLLOWING

WANDERING

= INHIBITION BETWEEN BEHAVIORS

(31)

Fly Vision

•Researchers at France’s

Neurocybernetics Research Group (CNRS) considered housefly’s

compound eye as a useful way for robot to view world

•Housefly’s visual navigation system consists of about 1,000,000 neurons

–Neurons constantly adjust amplitude, frequency, and twist of wings, which are controlled by 17 muscles –Visual motion used for coarse control

•Eye of housefly:

(32)

Fly Vision (con’t.)

•

CNRS developed reactive mobile robot that uses insect-like visual system

(33)

Fly Vision (con’t.)

•

Biological principles exploited include:

–Use of compound optic design, generating a panoramic view

–Visuomotor control conducted using optical flow induced by robot’s motion –Locomotion consisting of a succession of translational movements followed

by abrupt rotations, typical of fly’s free-flight behavior

–Motion detection circuitry based on electrophysiological analysis of the housefly using analog design

–Use of space-preserving topographic (retinotopic) mappings onto the control system

–Modeling from an invertebrate perspective, using an exoskeleton as

(34)

Ant Chemotaxis

•

Ants: relatively simple creatures capable of complex

actions through their social behavior and their interactions with the environment

•

Ant communication:

–Predominantly chemical

–Visited paths marked with volatile trail pheromone

–Ants traveling path continually add to odor trail, strengthening it

•

Many researchers studying this mechanism in simulation

–Biologically plausible trail generation using mathematical behavior models –Production of species-specific foraging patterns for 3 species

(35)

Ant Chemotaxis (con’t.)

• Researchers in Australia (Russell, Thiel, Mackay-Sim): created robot systems capable of laying down and detecting chemical trails

• Camphor serves as chemical scent

• Depositing trail:

–Robot drags a felt-tipped pen containing camphor across the floor as it moves, depositing trail 1 cm wide

• Sensing trail:

–Two sensor heads separated by 50 mm

–Inlet draws in air from immediately below sensor across a gravimetric detector crystal –Detector crystal treated with coating that absorbs camphor

–As mass added, crystal’s resonant frequency changes in proportion to amount of camphor absorbed

(36)

Summary of Principles and Issues

in Transferring Biological Insights to Robots

• Programs should decompose complex actions into independent behaviors, which tightly couple sensing and acting. Behaviors are inherently parallel and distributed.

• To simplify control and coordination of behaviors, agent should use straightforward, boolean activation mechanism (e.g., IRM)

• To simplify sensing, perception should filter sensing and consider only what is relevant to the behavior (I.e., action- oriented perception)

• Direct perception (affordances) reduces the computational complexity of sensing

• Behaviors are independent, but the output from one may be combined with another to produce a resultant output, or may serve to inhibit another

(37)

Classical robot behaviors

Exploration/directional behaviors (move in a general direction)

wandering

Goal-oriented appetitive behaviors (move towards and attractor)

discrete object attractor area attractor

(38)

Classical robot behaviors

Exploration/directional behaviors (move in a general direction)

wandering

Goal-oriented appetitive behaviors (move towards and attractor)

discrete object attractor area attractor

(39)

Aversive/protective behaviors (prevent collision)

avoid stationary objects

elude moving object (dodge, escape) aggression

Path following behaviors (move on a designated path

road following

hallway navigation

(40)

Postutal behaviors

balance stability

Social/cooperative behaviors

sharing foraging

flocking/herding

(41)

Teleautonomous behavior (coordinate with humann operator)

influence

behavioral modification

Perceptual behaviors

saccades

visual search ocular reflexes

(42)

Walking behavior (for legged robots)

gait control

Manipulation specific behaviors (for arm control)

reaching

gripper/dexterous hand behaviors (for object)

grasping enveloping