Some evidence for predictive control of the capture of moving targets by the echolocating bat, Eptes

1
Some evidence for predictive control of the
capture of moving targets by the echolocating
bat, Eptesicus fuscus

• Harry R. Erwin, Ph.D., The School of Computing,
• Information and Technology, University of
  Sunderland,
• Sunderland, SR6 0DD, United Kingdom
• Cynthia F. Moss, Ph.D., Department of Psychology
• Program in Neuroscience and Cognitive Science
• University of Maryland
• College Park, Maryland 20742, USA

2
Abstract

• A computational sensorimotor model of target
  capture behavior by the echolocating bat,
  Eptesicus fuscus, has been developed that
  integrates acoustics, target localization
  processes, flight aerodynamics, and target
  capture planning to produce model trajectories
  replicating those observed in behavioral trials.
  
• Using this model, we have explored the roles of
  the approach and terminal phases of target
  capture in FM-bats. Exercise of this model
  against data from behavioral trials with moving
  targets in circular motion has provided evidence
  that bats use predictive tracking with a control
  algorithm that predicts and adjusts for
  non-constant target acceleration in addition to
  target motion and position.
• An echo state machine model based on the ideas
  of Jaeger, Maass, Natschlaeger, and Markham is
  currently being explored to see if it provides
  insight into possible mechanisms for doing this
  prediction.

3
The basic problem how the bat captures prey
using echolocation

• Figure from Webster and Brazier, Experimental
  Studies on Target Detection, Evaluation and
  Interception by Echo-locating Bats , 1965.
• A bat (Myotis lucifugus) capturing a moth in
  foliage.
• 100 millisecond intervals.
• The bat had first detected the tree about 500
  milliseconds before the first image.
• Data available to the bata few biosonar
  snapshots in the dark.

4
Research goal To understand sensory-motor
integration in bat

• Using static targets, we first examined the
  following questions
• What localization cues does the bat use?
• What flight control algorithm does the bat use?
• What is the aerodynamic behavior of the bat?
• Using moving targets, we now addressed
• Is the capture control algorithm predictive or
  non-predictive? Non-predictive control would be
  expected to lag a moving target.

5
Predictive versus non-predictive

• Predictive use models of the targets motion and
  of the bats self-motion to select a capture
  strategy to optimize the capture probability.
• Non-Predictive use the current state estimate or
  last known target localization to control the
  capture.
• Simple homing.
• Lead pursuit.
• Lag pursuit.

6
A gray-box model of bat echolocation and target
capture

• Founded on work (Kuc 95) in sonar-controlled
  robotics, but more biologically realistic.
• Realistic aerodynamic models.
• Calibrated from behavioral trial data.
• Model behavior based on a world model updated
  asynchronously by the echolocation of targets of
  interest.

7
Behavioral trials

• Measurements of the flight and acoustic behavior
  of bats during capture of tethered mealworms,
  Tenebrio molitor, in an open room were used to
  calibrate and validate the model.
• The original model was calibrated based on 16
  behavioral trials and validated against 15
  additional trials with stationary targets.
• The model was then exercised against 13
  behavioral trials with targets moving in circular
  motion.

8
Behavioral data collection

• Two genlocked (frame synchronized), high-speed
  video cameras (Kodak MotionCorder, 640 x 240
  pixels, 240 Hz frame rate, and 1/240 second
  shutter speed) were positioned just below the
  ceiling in opposite corners of the flight room.
• A volume 2.2 m x 2.2 m (horizontal) x 1.6 m
  (vertical) within the region defined a calibrated
  space for reliable 3-D reconstruction of the
  bats flight path.
• A calibration frame (Peak Performance
  Technologies) was placed in the center of the
  flight room and video-taped by both cameras prior
  to each recording session.

9
Behavioral data processing

• A commercial motion analysis system (Peak
  Performance Technologies Motion Analysis
  System-Motus) was used to digitize both camera
  views using a Miro DC-30 Plus interface, to
  calculate the 3-D location of the bat and
  mealworm target.
• The calibration procedure used in the video
  processing produced a mean error of 1.0
  centimeter in each coordinate.
• The bats initial velocity was estimated from
  position measurements separated by a time
  interval of 100 msec, producing a mean error of
  0.245 m/sec.

10
Model performance with a stationary target
Top view of trajectories Scale in meters
11
Conclusions from stationary targets

• Localization makes use of range, azimuth and
  elevation cues.
• Capture behavior is more than simple homingthe
  bat seems to maneuver for a good position above
  the target to perform a homing capture.
• Non-predictive capture control an adequate model
  for stationary targets.
• Aerodynamics constrain target capture by limiting
  the points the bat can reach without maneuvering.
• Trajectories are sensitively dependent on when
  targets are detected and decisions made.

12
Non-predictive capture control algorithms examined

• Non-predictive capture control with the model bat
  flying to the last known location of the target.
• Semi-predictive capture control with the
  position and velocity of the target maintained
  using a Benedict-Bordner ??? filter (optimal for
  tracking some maneuvering targets). The bat
  model then flies to the estimated current
  position of the target. A number of values of ?
  must be studied, since the values are a
  compromise between good smoothing and rapid
  response to maneuvering targets.

13
Predictive capture control algorithms examined

• A Benedict-Bordner ??? filter was also used to
  predict the capture point (rather than just the
  current location).
• Four unscented Kalman filters (UKF)
• Cartesian coordinates, Singer model.
• Cartesian coordinates, circular maneuvers.
• Head-centered coordinates, Singer model.
• Head-centered coordinates, circular maneuvers.
• A simplified Segmenting Track Identifier (STI)
  model

14
Segmenting Track Identifier (STI) capture control
algorithm

• Uses the complete set of target measurements to
  reestimate the parameters of the targets
  circling motion by a non-linear least-squares fit
  at each measurement update.
• If the algorithm fails to converge or if the
  resulting circle is unrealistic in position or
  size, a fit of linear motion to the target data
  is used.
• This corresponds conceptually to a biological
  model where the bat keeps the 4-dimensional
  trajectory and the supporting measurements in
  short term memory and predicts ahead based on
  past experience.

15
Actual performance of a bat capturing a moving
target
Top view Scale in meters
16
Performance of a non-predictive algorithm with a
moving target
Top view Velocity not being estimated. Results
suggest that the bat was predicting the targets
motion to the capture point.
17
Performance improvement using a Benedict-Bordner
??? tracking filter
Top view Using location and velocity estimates
to predict the capture position. This approach
generally performed well but usually not as well
as shown here. Poorer than behavioral performance.
18
Some evidence for tracking in accelerationside
view
Length of acoustic recording 8 sec. Length of
video recording 2.7 sec. Markers indicate the
bat and target locations at the times when the
bat vocalized. Arrows show bat motion. Target in
circular motion.
19
Evidence for tracking in accelerationtop view
Markers are at positions when the bat
vocalized. Arrows show motion of the target and
bat. The bat flew a complete circle before
coming back to make the capture.
20
Performance of predictive ??? capture control
algorithm
The large minimum distance to target seen here
(about 0.1 m) was typical for all values of a
explored (0.1-0.7). This suggests that the real
bat was taking account of target acceleration in
controlling its flight.
21
Performance of a Kalman filter
The performance of the various Kalman filters
studied was inferior to both the ??? and STI
filters. This was not surprisingsonar
measurements are nonlinearly related to the
target state, and so Kalman filters generally
have problems with sonar tracking.
22
Performance of the STI filter
The performance of the segmenting track
identifier model approached that of the real bat.
The distance of nearest approach was less than
3.9 cm.
23
Summary of distance of nearest approach (13
behavioral trials)
24
Modeling trajectory prediction using neural
microcircuits

• Being performed at the University of Sunderland
• References
• H. Jaeger and H. Haas, "Harnessing nonlinearity
  predicting chaotic systems and saving energy in
  wireless telecommunication," Science, vol. 304,
  pp. 78-80, 2004.
• W. Maass, T. Natschläger, and H. Markram,
  "Real-time computing without stable states a new
  framework for neural computation based on
  perturbations," Neural Computation, vol. 14, pp.
  2531-2560, 2002, and
• N. Bertschinger and T. Natschläger, "Real-time
  computation at the edge of chaos in recurrent
  neural networks," Neural Computation, vol. 16,
  pp. 1413-1436, 2004.

25
Background

• Natschläger, et al, 2002, suggest that the
  stereotypical neural microcircuits of the cortex
  may be the computational units of the brain.
• These microcircuits appear well-adapted to
  handling continuous streams of information, but
  sensorimotor integration in actively echolocating
  bats requires a computational unit that can
  generate a continuous output stream representing
  the location of a target from asynchronously
  received discrete echo returns.
• We are now investigating whether this model can
  be applied to this problem.

26
Modeling Approach

• Initially using an Echo State Machine (Jaeger).
• Target trajectory is represented by an array of
  place cells organized in general Cartesian
  coordinates. 20 msec time intervals.
• Sensory afference is structured similarly. Flag
  variable indicates a cry was received during an
  interval.
• Output trajectory reflects the echo reservoir,
  inputs, previous outputs, and any flag variables.
• First results suggest that distributed
  representations may work better than individual
  binary place cells.

27
Model Structure
Input Array
OutputArray
Echo State Reservoir
Circle indicates internal feedback
28
Conclusions

• Bats adjust for non-steady target acceleration
  and velocity, performing better than good Kalman
  filters.
• This suggests bats use recent target trajectory
  history and measurements to generate and test
  hypotheses about the future motion based on their
  experience with similar targets.
• A simple STI filter algorithm was designed to
  model this process, and was shown to produce
  performance approaching that of real bats in
  behavioral trials.
• Neural microcircuit models of this process are
  currently being investigated.

29
With Thanks To

• Peter Abrams, Myriam Tron, and Amy Kryjak
  assisted in the data collection and analysis of
  behavioral trials.
• Paul Kelley developed the custom software used to
  trim the digitized audio data.
• Aaron Schurger produced 3-D visualizations of the
  trial and model trajectories.
• This research extended behavioral studies
  performed by Willard W. Wilson, Ph. D.