Biologically-inspired
visual guidance of robot locomotion
[see Simon K. Rushton
, Jia Wen & Robert S.
Allison . Egocentric Direction and the Visual Guidance of Robot
Locomotion Background, Theory and Implementation. BMCV (forthcoming) .]
Overview
Drawing on
the recent suggestion [1] that humans rely on
perceived egocentric direction, rather than optic flow to guide
locomotion we implemented a robot guidance system.
The system
relies on some ideas outlined elsewhere [2] to
guide target interception of static and moving targets.
For an brief
overview of some of the experimental work that undepins this work,
please check Simon's research page
here .
The first
assumption of the egocentric model is that approach to a target is
based upon maintaining a constant egocentric direction. Simply
put, you can intercept a target if on each step you check that the
target is at its previous egocentric direction, and if it is not then
turn so as to fix this. Provided that the direction you are
keeping the target at is less than 90degrees (measured relative to your
locomotor axis), then you will reach your target. The path you
will take is an equi-angular spiral. Illustrated below are a
family of constant-eccentricity trajectories (plan views) that
intercept (1) a static target; (2) a target moving with a constant
velocity; (3) an accelerating target.
Sometimes a
system won't be calibrated (eg you don't know whether your target is at
15 degrees or -20 degrees), one way to deal with this is to use target
drift [3] . The image below shows the use of target
drift and over-compensation for guidance of an uncalibrated robot and
calibration.
We also
added an obstacle avoidance system that uses basic visual variables
such as time-to-contact (TTC; 4 , 5
), and trajectory [6 ,7 ].
Selection and optimisation of the calculation of these variables is
based upon recent experimental findings and computational models [ 8 ].
A simple
control law, derived from work on human interception [ 9
] and body-scaled parameters[10 ] can be used to
produce successful avoidance of static and moving obstacles.
Examples
note .avi files.
Moving
obstacles
Different
obstacle's
shapes - Triangle & Planar
Different
sizes
Interface
Our current
implementation includes a simple to use interface for testing, an
object-orientated Matlab implementation and the ability for batch
processing of trials for performance evaluation.
We also
have a partial implementation on a Nomad robot (lack of space to run
the robot
being the constraining factor).
This work
is in development. Further information is available on request. The
work will be presented at BMCV in November. Others [ 11
, 12 ] have done work on the same problem, using
different approaches. We believe our model has significant advantages.
Future
projects -
Computational
Refined
perception of direction
Perception
of direction with a mobile stereo head
High-level
situations-specific algorithms (such as recognition of dead-ends)
Full
implementation on Nomad robot (partial implementation already)*
Further
refinementy and optomisation of control laws
Further
refinement and optomisation of calculation or pick-up of visual
varaiables (such as
TTC)
More
complex
scene segmentation
Human
trajectory and eye-movements
Perception
of egocentric direction: disparity field
Perception
of direction: flow field
Investigation
of the influence of multiple obstacles
Selection
of
closest target by distance, TTC, direction, TTP?
Teleoperation
of visually guided robot
* Jia Wen,
4080 project, Winter term
If you are
interested in any of these projects please email
References
[1] Rushton, S.K., Harris, J.M., Lloyd, M.L. & Wann,
J.P. (1998). Guidance of locomotion on foot uses perceived target
location rather than optic flow. Current Biology, 8, 1191-1194.
[2] Rushton, S.K. & Harris, J.M. (submitted). The
utility of not changing direction and the visual guidance of locomotion
.
[3] Llewellyn
KR. (1971). Visual guidance of locomotion. Journal of Experimental
Psychology, 9, 1245-261.
[4] Lee, D.N.
(1976). A theory of visual control of braking based on information
about
time-to-collision. Perception, 5, 437-459.
[5] Regan, D.
& Hamstra, S. (1993). Dissociation of discrimination thresholds for
time
to contact and for rate of angular expansion. Vision Research,
33,
447–462
[6] Bootsma, R.J. (1991). Predictive information and the
control of action: what you see is what you get. International
Journal of Sports Psychology, 22, 271–278.
[7] Regan, D.
(1993). Binocular correlates of the direction of motion in depth. Vision
Research, 33, 2359-2360.
[8] Rushton, S.K. & Wann, J.P. (1999). Weighted
combination of size and disparity: a computational model for timing a
ball catch. Nature Neuroscience, 2, 186-190.
[9] Peper, L., Bootsma, R,J,, Mestre, D.R. & Bakker,
F.C. (1994). Catching Balls - How To Get The Hand To The Right Place At
The Right Time. Journal of Experimental Psychology-Human Perception
And Performance , 20. 591-612.
[10] Warren –
affordances and stepping
[11] Fajen, B.R., Warren, W.H., Temizer, S & Kaelbling,
L.P. (in press). A dynamical model of visually-guided steering,
obstacle avoidance, and route selection. International Journal of
Computer Vision
[12] Khatib, O. (1986). Real-time obstacle avoidance for
manipulators and mobile robots.
International Journal of Robotics Research, 5. 90-99.
Acknowledgements
This research
was supported in part by funds from National Science and Engineering
Research Council of Canada and Nissan Technical Center North America
Inc.