Robotic Electrolocation: Active Underwater Target Localization with Electric Fields

1
Robotic ElectrolocationActive Underwater Target
Localization with Electric Fields

• James R. Solberg Kevin M. Lynch
  Malcolm A. MacIver

2
Inspiration Weakly Electric Fish
Dielectric Object
E-field Flux
black ghost knifefish (Apteronotus albifrons)
Voltage Detectors

• Self-generated electric field (1 mV/cm near its
  skin)
• Detect perturbations in field due to nearby
  objects

3
Active Sensing in Weakly Electric Fish

• Weakly electric fish utilize both definitions of
  the term active sensing
• Active Sensing 1 Transduce self-generated
  energy.
• Active Sensing 2 Move sensors to gather better
  information.
• In this talk we investigate the latter
  definition.

4
2-DOF Robotic Active Electrolocator
x linear slide
generated electric field
voltage detectors
target
y linear slide
5
Possible Electrosense Application Close-Range
Sensing for AUVs

• Short range (lt 1m)
• High resolution (lt 100 µm)
• Omnidirectional
• Well suited for low-speed, highly-maneuverable
  AUVs

6
Overview

• Estimate target location from voltage
  measurements --- Electrolocation
• Implement an active control scheme on an XY robot
  with the task of electrolocation.

7
Simple 2-D Example
Define observation w V1 V2
V 0 volts
w 0 mV
V1
V2
8
Target in Electric Field
At x 20mm,-20mm, w -44 mV
Contours at 50mV
V1
V1
electrical conductor
V2
V2
V 0 volts
9
Observation, w, as a function of target location,
x
w -44mV
voltage detector
emitter
mV
emitter
x 20,-20, w -44 This is the data point
found from the previous slide.
voltage detector
10
Probabilistic Sensor Model, p(wx)

• For real data must explicitly account for sensor
  noise.
• Collect many ws at each x to construct empirical
  sensor model.
• Parameterize with mean, µw, and variance,
  sw2(noise is Gaussian), for all possible target
  locations, x

2-D slice
µw
10 V
sw
p(w x)
w likelihood for this x
w
x is the position of the target relative to the
center of the robot
µw as a function of x
-10 V
11
Detection Range
µw as a function of x

• Yellow contour is 95 confidence detection

55mm
45mm
12
Sensor Fusion via Particle Filter

• Represent belief of target location as set of
  particles.
• Each particle is a possible location of the
  target.
• Recursive Bayes filter updates the particles.
• The spatial variance of the particles is a
  measure of the uncertainty of the target location

region of highest probability
Each particle is an instance (hypothesis) in
state space
13
Active Control of Robot

• Controller objective Minimize expected particle
  spatial variance at next time step (greedy).
• Simulate each control option forward one time
  step.
• Choose control option that yields the expected
  belief with the lowest uncertainty
• Robot ends trial when belief (particles)
  uncertainty is sufficiently low. Our metric is
  the square root of the trace of the covariance
  matrix

Trial stopping condition
mm
(uncertainty metric)
14
Electrolocation Experiments

• XY robot moves to estimate location of spherical
  target
• 8 experimental conditions
• 2 diameters 12.7mm (½) and 38.1mm (1-½)
• 2 target materials metal plastic
• 2 water salinities fresh and salt
• No motion uncertainty
• Robot moves at constant z above the target.

15
Electrolocation Workspace

• Visits each grid point to construct sensor model
• Starting positions are randomly sampled from blue
  box
• Blue Gray box are permissible locations for the
  center of the robot
• Control options are randomly sampled from orange
  box.

detector
emitter
white is robot position
small target workspace
top view of workspace
large target workspace
emitter
detector
16
Active Controller in Electrolocation
Controller minimize expected particle spatial
variance at next time step
(1) Initial belief
(1) observation update
(2) move
(2) observation update
w2 4 V
w1 0 V
(3) observation update
(4) move
(4) observation update
(3) move
w4 -4 V
w3 -2 V
17
Final Belief of Trial
Centroid of 2000 particles (red x) is the
estimate of the target
Actual location of target is x0, y0.
Orange asterisks are the four locations visited
by the center of the robot
18
Choosing Next Prospective Greedy vs. Random

• Compare active controller with choosing control
  option randomly (random walk)
• For each of the 8 experimental conditions, 50
  electrolocation trials were performed for each of
  the two controllers (800 trials total)
• A trial has failed if either of the following
  conditions are met
• Trial takes more than 35 steps to locate the
  target
• - or -
• The difference between the estimate (centroid of
  particles) and actual position is greater than 15
  mm.

19
Comparison of Controllers
Number of failed trials (out of 50)
Fresh Water
Salt Water
metal plastic
Median number of steps for completion
20
Summary

• Implemented a sensing modality (inspired by
  weakly electric fish) capable of estimating the
  location of a target based on voltage
  measurements from a self-generated electric field
• This electrosense was used to control an XY robot
  to actively locate the position of the target.
• The active control scheme performed better than a
  random walk.