lect7

advertisement
Adaptive Robotics
COM2110
Autumn Semester 2008
Lecturer: Amanda Sharkey
1
Robots in the news








9:13am UK, Friday November 14, 2008
“British scientists have come up with the first robot that can
mimic a person's expressions simply by watching their face.”
A disembodied humanoid robotic head with video camera eyes
and 34 internal motors under “Flubber” skin
Works with 10 human emotions (happiness, sadness, concern)
Can mimic human expressions that it recognises.
Bristol Robotics Lab (Chris Melhuish) at UWE
Robot head commissioned from David Hanson
Can scan and “understand” images at 25 frames per second.
2








Lect 1: what is a robot? Brief history of robotics
Early robots, Shakey and GOFAI, Behaviour-based robotics
Mechanisms and robot control (and biological inspiration)
Lect 2: Grey Walter, Brooks and Subsumption Architecture.
Lect 3: Adaptation and learning
Lect 4: Artificial Neural Nets and Learning
Lect 5: Evolutionary Robotics
Lect 6: Swarm Robotics (reasons for, biological inspiration, local
control and communication, self-organisation and emergence).
3
Robots as Biological Models


Robots can be used to investigate
biological questions
And to test hypotheses
4
Biorobotic Modelling



What is Biorobotics?
The intersection of biology and robotics
Common ground between robots and
animals:



Both are moving, behaving, systems
Both have sensors and actuators
Both require an autonomous control system
to carry out tasks in a dynamic world.
5

To count as biorobotic modelling a system
should


Be robotic (physically instantiated)
Be biological (address a biological
hypothesis, or demonstrate understanding of a
biological system). (Webb, 2001)
6
Biorobotics:
“understanding biology to build robots and
building robots to understand biology”
(Webb, 2001)
7
Synthetic modelling

Understanding by building

Synthetic vs analytic


Analytic approach in empirical sciences –
performing experiments on existing systems (human,
ant, brain region) and analysing results
Synthetic: Idea of creating an artificial system that
reproduces certain aspects of natural system
8

Robots have (arguably) replaced computers
as the metaphor to develop and test models
of biological intelligence
9




Early example – “electric dog” (Seleno)
Loeb and animal taxis or forced movement
Hammond and Miessner,(1912) “electric
dog”
Loeb saw the “artificial heliotrope
machine” as confirmation of his theories
10
Advantages of biorobotic models




Robotic implementations force scientists to be
very concrete in specifying the design of
biological system
Robot operating in the real world constrains the
choice of models
Can help to produce testable hypotheses
Robot models allow scientists to study interaction
of model with the environment

Simulations may introduce misleading
simplifications
11
Webb (2001)

1.
2.
3.
4.
5.
6.
7.
7 dimensions along which robotic models of biology differ
Relevance: whether model tests and generates hypotheses applicable
to biology
Level: the elemental units (atoms – society)
Generality: the range of biological systems the model can represent
Abstraction: the complexity relative to the original
Structural accuracy: how well the model represents the actual
mechanisms underlying behaviour
Performance match: to what extent the model behaviour matches the
target behaviour
Medium: the physical basis by which the model is implemented.
12
A robotic cricket: Song Recognition
and Localisation




Lund, Webb and Hallam (1998)
Phonotaxis in the cricket
Lund, H.H., Webb, B., and Hallam, J. (1998) Physical and
Temporal Scaling Considerations in a Robot Model of
Cricket Calling Song Preference. Artificial Life, 4, 1, 95107.
Webb, B., and Scutt, T. (2000) A simple latencydependent spiking-neuron model of cricket phonotaxis.
Biological Cybernetics, 82: 247-269.
13




Female crickets recognise males’ songs and
approach them.
They have an ear in each foreleg that
produces direction-dependent differences
in response amplitude
Turns to strongest response?
But – only approach males of the same
species with characteristic song.
14
Field cricket
15
Ground cricket
16
Tree cricket
17
Mole cricket
18
Bush cricket
19
Scaly cricket
20





Taxis, or approach to a sensory source
When modelled, often abstracted as a source
signal, whose value decreases with distance, and
an animat that can sense that value and use it to
control movement to the source.
E.g Braitenberg’s vehicle 2
But does this capture the essence of the
behaviour?
Lund et al (1998) argue it does NOT
21
They say …


Abstract model ignores the real physics of signal
propagation and detection
Reasonable approximation of light but not of
other modalities



E.g. chemotaxis, where odour signal is highly
dispersed and carried by currents in environment
Phonotaxis: sensors don’t respond exclusively to
relevant signal
Emphasis on physical properties of task and agent
22
Unidirectional communication
- Male cricket stridulates
-
-
Rubbing wings together, or a leg against a
wing.
Female cricket responds by moving
towards the source of the song.
23
24


Earlier implementation: LEGO robot – but
slow and comparison to cricket data
difficult.
Here, new robot and revised models


band-pass selectivity
using signals that are temporally identical to
crickets.
25




The female cricket has 4 auditory openings
An ear (tympanum) on each foreleg
An auditory spinacle (hole) on each side of
the frontal section
The 4 openings are linked by tracheal tubes


Sound reaches tympani through air
And via internal tubes from other auditory
openings
26
27




Sounds transduced from tympani are a
combination of delayed and filtered signals
The delays and filters improve cricket’s ability to
discriminate the arrival direction of its mate’s
song.
Sounds arriving from same side as tympanum are
delayed (by internal structure) to arrive in
antiphase to ipsilateral ear, and in phase with
contralateral ear.
Sounds subtracted, so intensity at ipsilateral
(same side) ear is enhanced
28
Female crickets home in on potential partners by
listening to their chirping (phonotaxis)
Sound is made of short bursts or syllables.
She moves towards the source of the song;
But she only responds to calling songs with syllable rates
within a certain bandpass.
Are two processes implicated?
(i) recognition of cricket song
(ii) localisation system to approach source of the
song.
29
30
31
32
The neural model





Input from auditory sensors is fed into a neuron on
each side (N1)
Activation can flow from N1 to another neuron N2
that feeds activation to the motor on that side.
Activation modelled with a leaky integrator with
associated threshold (T high) before firing.
When it is fired, activation has to decay below a
lower threshold (T low) before it can fire again
NB N1 and N2 are not intended to correspond to
specific neurons in the cricket, but represent
processes carried out by small numbers of neurons.
33
34
Phonotaxis experimental results
The robot is attracted to the male cricket calling song. It
also discriminates between calling songs with the right
carrier frequency (4.7kHz) and those with other carrier
frequences (e.g. 6.7kHz)
Neural circuitry is set up to respond to species-specific
syllable repetition interval
35
36
37

The robot’s response to different syllable
rates (little response when syllable intervals
are less than 20 ms and more than 60 ms) is
the same as in the real cricket.
38
Their conclusions




Explains why crickets respond only to certain calls
Mechanism implements the bandpass filtering found in
crickets.
With N1 neurons acting as low pass filter, and N2 acting
as high pass filter.
“high level” model of cricket controller – N1 and N2
don’t correspond to specific neurons in cricket, but
represent processes that may be carried out by 3-10
neurons in the cricket prothoracic ganglion and brain.
39


Mechanism underlying phonotaxis in
robotic cricket is surprisingly simple
Shows that there is not need for two neural
control systems



One to recognise the call
One to locate it.
Here the two systems are one
40

“A generic simulation may tell us little
about real problems in approach behaviour.
By investigating a specific biological
system and modelling it at a level of detail
driven by biological questions, we gain
more sophisticated insights into the real
problems of sensorimotor control”

Lund et al (1998)
41
But it is also possible to investigate phonotaxis in real
crickets …..
42






Hedwig, B., and Poulet, J. (2004) in Nature
Using tracking ball, could look at cricket
responses to chirps presented from left and right
of cricket
Response was very rapid
- “reflex like reactive steering towards individual
sound pulses”
Other methods (bandpass filtering brain neurons,
or template matching), would require at least two
consecutive sound pulses
But recognition process may modulate steering
responses over time.
43
Biorobotics and ….
Biological relevance: biorobotics more likely to
assume a particular biological target than
biologically inspired robotics
 Level: usually sensor transductions considered
 Generality: usually biorobotic models are more
specific e.g modelling sound localisation of
crickets. But general principles can emerge (I.e.
idea of exploiting timing properties of neural
firing)

44


Abstraction. Usually modelling simplifies some
aspects – e.g. in the robot cricket model there is
minimal representation of biological details in
motor control of cricket
Structural accuracy: how well do mechanisms
reflect real mechanisms in target? How would
we know?
Usual aim is to build complete model that
connects sensing to acting.
45

Performance Match: to what extent does model
behave like the target?
How will model’s performance be assessed?
Little possibility of falsifying model
- if match is poor this could be attributed to
measurements, not model
- depends on interpretation of model behaviour
- parameters of model can be tuned until they
match.
46

Medium: what is simulation built from?

Biorobotics, emphasis on “complete model” that
produces an output, and on understanding the
environmental conditions
47
What does a biorobotic model tell
us?


Can we ever really validate a model?
Do models ever tell us something we didn’t
already know?
48
Modelling: involves the correspondence between a
real target system and something else.
Problems: underdetermination. There are many
possible mechanisms that could underlie a
behaviour. If two systems behave the same, it
does not follow that the cause or mechanism of
the behaviour is the same.
But, can show that a proposed mechanism is
sufficient to produce the behaviour
49
Can we assume that this model of cricket
phonotaxis represents the mechanism
underlying the cricket’s behaviour?
Reasons for doubting:
50
Cataglyphus – the desert ant
51

How does the desert ant return to nest after
foraging?


Cannot use pheromones – they evaporate too
quickly
It uses a combination of strategies
Path integration
 Visual piloting
 Systematic search

52


Main strategy: path integration
Uses compass information based on
polarization pattern of the sky


Directions of polarization (e-vectors) form
regular patterns, as sunlight is scattered by
atmospheric molecules.
Polarization pattern of sky is invisible to
humans, but can be exploited by insects
53
Path Integration
Random search for food, but almost straight path home when food is found
54
The desert environment with few landmarks
55
Lambrinos and his colleagues built a navigation system for Sahabot,
based on the ant’s solar compass.
Lambrinos, D. et al (1997)
56
The compound eye of
the ant is made up of
many ommitidia. It is the
upper ommitidia the
house the solar compass.
57
Polarised light sensors
sensors
58
59
Sahabot


Could use robot to test out different models
of acquiring compass information from the
polarized light pattern of the sky
And to test out models of visual landmark
navigation
60



Robot tested in the same environment as
ants
But different in size, and method of
propulsion
Experiments provide confirmatory
evidence for biological hypotheses
61





Lambrinos et al (1997, 2000)
Goal of synthetic methodology, “to develop an
understanding of natural systems by building a
robot that mimics some aspects of their sensory
and nervous systems and their behaviour”
Biorobotic study can result in innovative
engineering designs
And confirm biological hypotheses
And generate new biological hypotheses
62
“This case study illustrates the power of biorobotics
that arises from the close relationship between
engineering and biology. On the one hand,
insights for innovative engineering designs can be
found by analyzing the mechanisms employed by
biological agents. On the other hand, biological
hypotheses can be confirmed using real world
artifacts rather than simulation only, and new
biological hypotheses and ideas for new animal
experiments can be generated” (Lambrinos et al,
2000).
63
Task allocation

Krieger, M.J.B, and Billeter, J-B (2000) The call of duty:
Self-organised task allocation in a population of up to
twelve mobile robots. Robotics and Autonomous
Systems, 30, 65-84.

How are tasks allocated in insect colonies?
- different castes? (genetic)
Or activation thresholds?
Relevancy to engineering applications?



64
Activation thresholds



Reaction to stimuli
E.g. neglected brood, or corpses of dead
ants – odour increases in strength
When it reaches an individual’s threshold
value, they respond

E.g. grooming brood, or carrying corpse out
of nest.
65


Use of activation thresholds for
coordination of robot team
- aim to show that implementing different
activation thresholds is sufficient for
effective task allocation


Join waiting line
Exploration of a possible mechanism for
self-organisation of social insect societies
66
67
68
69


Robots’ mission – to collect food items and
bring them back to nest to keep nest energy
at safe level
Two tasks


Remain in nest
Forage
70

Wait in nest



Leave nest


Leave nest when nest energy lower than activation
threshold
(knowledge of nest energy)
No knowledge of nest energy
Look for food items




Load food item
Return to nest
Radio in energy consumption and recharge personal
energy store
Unload food item
71
Experiments


Nest energy remained stable – for teams of
3,6 and 9 robots. Decreased for teams of
12 robots.
Experimental manipulations:


Group size and food distribution
See Figure
72
73

Advantages of activation threshold:


Decentralised method, can be applied to
variable sized groups
Method: equal distribution of activation
thresholds between lower and upper value

Effect of other methods? (not explored)
74

Conclusions of Krieger and Billeter (2000)

Individual activation thresholds are an
effective way of allocating tasks



Simple mechanism shown to be sufficient
Complex social systems can be regulated in
decentralised way
Groups of 3-6 robots did better than single
robots

Illustrates advantages of transition to social group
75
Summary

What is biorobotics?



Contrast between



Biological inspiration
Biological modelling
Examples of biorobotic models:




The intersection between biology and robotics
Robotic model that tests biological hypotheses
Model of cricket phonotaxis (Lund et al 1998)
Study of navigation in desert ant (Lambrinos et al 2000)
Study of task allocation (Krieger and Billeter, 2000)
Question: What can biorobotic models tell us?
76

What can biorobotic models tell us?




Explicit form of modelling – dealing with real
environmental stimuli
Can show that mechanism is sufficient to account for
behaviour
But cannot be sure that the model is the mechanism
used in biology (under-determination)
Often the challenge is to find simplest mechanism
77
Download