Research in the domain of biologically inspired walking machines has focused
for the most part on the mechanical designs and locomotion control.
Although some of this research has been concentrated on the generation of
a reactive behavior of walking machines, it has been restricted only to a few
of such reactive behaviors. However, from this research, there are only few
examples where different behaviors have been implemented in one machine
at the same time. In general, these walking machines were solely designed
for pure locomotion, i.e. without sensing environmental stimuli.
Therefore, in this thesis, biologically inspired walking machines with different
reactive behaviors are presented. Inspired by obstacle avoidance and
escape behavior of scorpions and cockroaches, such behavior is implemented
in the walking machines as a negative tropism. On the other hand, a sound
induced behavior called “sound tropism”, in analogy to the prey capture behavior
of spiders, is employed as a model of a positive tropism. The biological
sensing systems which those animals use to trigger the described behaviors
are investigated so that they can be reproduced in the abstract form with
respect to their principle functionalities. In addition, the morphologies of
a salamander and a cockroach which are designed for efficient locomotion
are also taken into account for the leg and trunk designs of the four- and
six-legged walking machines, respectively.
Different behavior controls for generating the biologically inspired reactive
behaviors are developed on the basis of a modular neural structure. Each
behavior control consists of a neural preprocessing module and a neural control
module. Preprocessing is for sensory signals while the neural control
generates basic locomotion and changes the appropriate motions, e.g. turning
left, right or walking backward, with respect to sensory signals. Neural
preprocessing and control are formed by realizing discrete-time dynamical
properties of recurrent neural networks. Parts of the networks are generated
and optimized by using an evolutionary algorithm. Utilizing the modular
neural structure, the coupling of the neural control module with different
neural preprocessing modules leads to the desired behavior controllers, e.g.
obstacle avoidance and sound tropism. Furthermore, these behavior controllers
are then fused by using a sensor fusion technique consisting of lookup
table and time scheduling methods to obtain an effective behavior fusion
controller, whereby different neural preprocessing modules have to cooperate.
Eventually, all of these reactive behavior controllers together with the
physical sensor systems are implemented on the physical walking machines
to be tested in a real world environment. The fully equipped walking machines
can be seen as artificial perception-action systems. As a result, the
walking machine(s) is able to respond to environmental stimuli, e.g. wandering
around, sound tropism (positive tropism), avoiding obstacles and even
escaping from corners as well as deadlock situations (negative tropism). The
developed controller is universal in the sense that it can be implemented
on different types of walking machines, e.g. four- and six-legged walking
machines, giving comparably good results without changing parameters.