The construction of an artificial moth (AMOTH) Dr. Anna-Carin Backman, SLU Dr. Qian Han, SLU Dr. Niels Skals, SLU Mr. Kwok Chong, ULEIC Ms. Jing Gu, ULEIC Mr. Philipp Knuesel, INI Ms. Josie Mackenzie, ULEIC Mr. Marcus Sjoholm , SLU
Paul F.M.J. Verschure Institute of Neuroinformatics University/ETH Zürich Switzerland
[email protected] www.ini.unizh.ch/~pfmjv www.amoth.org
Eric Chanie Alpha MOS
Bill Hansson Philipp Knuesel Paul Verschure Agr. University Sergi Bermudez Tim Pearce Alnarp INI Univ|ETH Zurich University of Leicester
Pawel Pyk INI Miki Carlsson Alnarp
Overview • • • • •
Moths and machines Why the moth? Sensor coding Behavioral control BBQ test
Research Artificial MOTH AMOTH Stainless Steel Rat
Ada
The brain
The Distributed Adaptive Control (DAC) architecture ! "
$% &
'
&
#
#
#
!
( (
( ) *
* *
AMOTH goals: Autonomous embedded exploration and monitoring
Why the moth pheromone system?
Olfactory search is opto-motor anemotactic
Courtesy of J. Hildebrand
Ishida & Morrizumi, Handbook of Machine Olfaction, Pearce, Schiffman, Nagle, Gardner,(eds.) Wiley-VCH, 2002.
Antenna and Receptors Sensilla
• • •
ORN: olfactory receptor neuron Odors enter small cavity in sensilla through pores ORNs show different degree of specialization Hansson, Trends Neurosci, 2002
Antennal Lobe
•
•
Spheroidal structure composed of densely packed neuropil called glomeruli Macroglomerular complex (MGC) and usually 50 to 100 ordinary glomeruli ORNs of same receptor type usually converge onto the same glomerulus MGC is specialized pheromone structure
•
Manduca sexta: ~66 GL, 360 LN, 900 PN
• •
Hansson, Trends Neurosci, 2002
Distler and Boeckh, J Comp Neurol, 1997
The moth can respond to concentrations of pheromones as low as 10-18g using 10-7g at its periphery
SNR= n Only 105 olfactory receptor neurons…..
How to model the moth‘s brain?
Ito et al., 1998
What’s the olfactory code?
Important compounds in coffee (E)-ß-Damascenone 1.95x10-1 2-Furfurylthiol 1.08 3-Mercapto- 3-methylbutylformate 3-Methyl-2-buten-1-thiol 2-Isobutyl-3-methoxypyrazine 8.30x10-2 5-Ethyl-4-hydroxy2-methyl-3(2H)-furanone Guaiacol 4.20 2,3-Butanedione (diacetyl) 4-Vinylguaiacol 6.48x101 2,3-Pentanedione 3.96x101 Methional 2.40x10-1 2-Isopropyl-3-methoxypyrazine Vanillin 4.80 4-Hydroxy-2,5-dimethyl3(2H)-furanone (Furaneol)` 1.09x102 2-Ethyl-3,5-dimethylpyrazine 3.30x10-1 2,3-Diethyl-5-methylpyrazine 9.50x10-2 3-Hydroxy-4,5-dimethyl2(5H)-furanone (Sotolon) 4-Ethylguaiacol 1.63 5-Ethyl-3-hydroxy-4-methyl2(5H)-furanone (Abhexon)
2.60x105 1.10x105 1.30x10-1 8.20x10-3 1.70x104
honey-like, fruity roasty (coffee) 3.70x104 catty, roasty 2.70x104 amine-like earthy
1.73x101 1.10x104 5.08x101 3.20x103 1.30x103 1.20x103 3.30x10-3 1.90x102
1.50x104 phenolic, spicy 3.40x103 buttery spicy buttery potato-like, sweet 8.30x102 earthy, roasty vanilla
1.70x103 1.70x102 1.00x102
caramel-like earthy, roasty earthy, roasty
1.47 7.50x101 3.00x101
seasoning-like spicy
1.60x10-1
2.00x101
Taken from: http://www.coffeeresearch.org/science/aromamain.htm Clarke, R. J. The Flavour of Coffee. In Dev. Food Science. 3 B. 1986. 1-47.
seasoning-like
Optical imaging of the glomerular response Camera Dye
PC
•Injection of dye •(FURA-dextran) •Imaged area •McGuire et al. (2001)
Glomerular response Benzaldehyde
Carlson et al (2005) Eur. J. Neurosci
E2-Hexenal
Odour encoding in the honeybee antennal lobe
Galizia et al. 1999
Quantifying the AL response
Kanzaki et al., 2003
Extracting the response parameters: Amplitude and Duration α(t; S;B;τ ; t0) = B + S (t-t0)/τ exp (- (t-t0)/τ) S: Amplitude B: Baseline τ: time constant t0: response onset.
Phenylacetaldehyde-PAA (plant compound)
Single versus combinatorial spatiotemporal coding
Temporal coding could play decisive role in sensitivity boost
Optomotor flight control
Insect Visual System: Structure
Model: Structure - Models the important processes involved in visual navigation - Two parallel paths that extract different features - Both neural models interact directly with the motor system via excitatory and inhibitory connections, which define the different priority levels
Collision Avoidance: LGMD Model
- Lobula Giant Movement Detector - Sensitive to Looming Stimuli - Receives input from on-off sensitive cells - Outcome = spikes
- δ = delay - X = Multiplication Locust lobula giant movement detector (LGMD) Image Claire Rind
LGMD physiology
A = vel exp(-α size)
Gabianni et al (2002) Nature
Model LGMD physiology A = vel exp(-α size)
Blimp Robot Tests - 2 wireless CCD cameras - Blimp based robot controlled via bluetooth communication - Combined collision avoidance model with self-stabilization model -
neural simulator software
Stabilization & Drift Compensation: Test Room Area – Top View - 6 test flights - Unbiased motor control
9.4°
- 7.1º mean deviation
15.6°
- 15.6 º max off course deviation - 0.62 m/s mean velocity
m.
- Compensation 59.5% time
Origin m.
Stabilization: Biased Motor Control (right motor @ 50%) Test Room Area – Top View
- 6 test flights - Biased motor control - 7.2º mean deviation - 10.7º max off course deviation - 0.23 m/s mean velocity - Compensation 66.3% time
m.
Collision Avoidance: Test Room Area – Top View
m.
end - Mean collision detection distance: 1.676 m. [1-2.7 m.] - Mean avoiding reaction time: 3.97 s m.
- 4 minutes flight
Chemical search
Chemosensor technology
Chemosensor board containing a 6 sensor array. The sensor sampling rate is around 1045 Hz. Then samples are averaged 64 times, which gives effective sampling frequency 16.3 Hz for each of the 6 sensor channels. Apart of serving the onboard sensor, the board has a connector with 2 additional analog inputs that allow for the integration of additional sensors, e.g. PID sensor and/or EAG. Weight: 13.8g, dimensions: 34.5x60 mm.
Indoor chemical search experiments
Indoor chemical search experiment
Wind tunnel calibration using static measurements of ethanol. (a) Response to a 9.4% solution of ethanol in distilled water. (b) Response to distilled water only (control).
The wind direction is the positive x-axis direction. The air flow in the wind tunnel was 1.097 m3 s-1 with an average air speed of 0.667 m s-1
Optomotor anemotactic chemical search experiment
Outdoor experiments
Outdoor chemical mapping Ethanol source 30
0
20
−5
20
15
−10
10
10
−15 0
y (m)
−20
y (m)
y (m)
5
0
−25
−10
−5
−30
−20
−35
−10
−40
−15
−45 −15
−10
−5
0
5 x (m)
10
15
20
25
−20
−30
−30
−25
−20
−15 −10 x (m)
−5
0
5
−15
−10
−5
0
5 x (m)
10
15
3 maps of the chemosensor response in a 40x40 m area derived from free flight trajectories
20
25
Conclusions • The moth is an excellent example for the construction of autonomous chemical search and mapping systems • Insect flight control systems can be successfully adapted to UAVs • Robot technology can be used to understand brain and behavior • Natural stimulus encoding strategies can be a great technological value • Monitoring and information gathering will be achieved by networks of autonomous perceptive and cognitive natural and artificial agents
Chemo-sensors: thin-film sensor technology • Low power consumption; • Sensor array miniaturisation; • Selectivity optimisation with geometry effect, dopant variations, change in operating conditions.
Example sensor response response to aprox 10 ml of cognac (sensor "looking" into a large cognac glass, aprox 10cm from beverage) sensor A; 68.75% x 2.5V = 1.71V
sensor B; 68.75% x 2.5V = 1.71V
sensor C; 68.75% x 2.5V = 1.71V
700
600
response
500
400
300
200
100
0 0
20
40
60
80 [sek]
100
120
140
160
publicity and media exposure Flying robot performance (Luzern, October 2004)
Sachsenleuten parade (March 2005)
SF DRS 1 Schweiz Aktuel October 2004
SF DRS 1 Kulturplatz (March 2005)
Physics for Kids (Zurich, March 2005)
Festival des Wissens (Zurich, March 2005)
cUAV autonomous control layer