Item

Abstract

Experimental results from insect biology suggest that in flies visual cues provide
important information for spatial orientation and flight control. Götz [1] suggested a
model that uses visual motion detection for course and altitude stabilisation. Similar
models can be used to explain obstacle avoidance behavior. Huber, Mallot and Bülthoff
[4] demonstrated that these biological principles can be applied to artificial systems,
and that even with a simple control architecture 2D obstacle avoidance behavior can be
achieved in a simulated autonomous agent.
We extend this approach to 3D and present a simulated flying autonomous agent that
uses only two elementary correlation-type local motion detectors [2] on each side for
horizontal and vertical motion, respectively. The agent’s head, containing all visual
receptors, is fixed with respect to the body coordinate system. The sensorimotor coupling
is provided by a simple feed-forward neural network from the motion detectors
to the motor system. In order to achieve 3D obstacle avoidance and flight stabilisation
behavior, the weighted connections are adjusted by a genetic algorithm in a closed
action-perception loop. The fitness values of the agents are determined from their performance
during an autonomous flight through a 3D virtual environment with obstacles
and simulated gravity. As in the original experiments with real flies, we use a sinusoidal
pattern for the simulated environment.
Simulation results show that 3D orientation and obstacle avoidance behavior is possible
with a simple control architecture. The agent evolves effective strategies for horizontal
and vertical obstacle avoidance, course stabilisation and altitude control. Qualitatively,
the weighted sensorimotor connections correspond with those predicted by Götz, i.e.,
they have the same sign. Simple exploration strategies in the environment can be observed
that resemble real fly behavior. For a successful evolution of 3D flight behavior,
the rotational motion of the agent has to be restricted to heading changes about the vertical
axis. Without this restriction, the agent would have to perform coordinate transformations
between the body and world coordinate systems in order to align the visually
perceived information with its attitude within the environment. This would require a
more complex information processing architecture. The rotation of the sensory system
can be restricted to heading changes by a separate roll and pitch stabilisation mechanism
for the agent’s head. Interestingly, real flies always keep their visual receptors
in an upright orientation with respect to the world coordinate system by mechanically
tilting their head up to 90 degrees during curved flight [3].