Targeting Behaviour in Cockroach Combat Modeled in a Robot

Heather Bell, Terry Stewart, Sergio Pellis, Ulysses Bernardet

The overwhelming majority of animal (and human) behavioural research is focused on identifying quantifiable elements of behaviour, and then measuring those elements. This approach comes with several problems. First, it assumes that what is measurable is what is relevant, and second, it assumes that behaviour is fundamentally stimulus-response (i.e., that it is unidirectionally causal; that it is able to be meaningfully parsed). Applying a cybernetic approach -- viewing animals as goal-directed control systems -- along with eliminating such problems, also brings the study of behaviour in line with what is known about the underlying physiology of organisms. In that vein, the aim of my project was to embody in robots a cybernetic rule purportedly being used during combat in Madagascar hissing cockroaches to see if it is sufficient to explain the variety of "unique" behaviours exhibited during bouts. Specifically, the goal of male roaches during combat is to flip their opponent, which is best achieved when the attacking male faces the flank of the defender at a 90 degrees. Since both animals have the same goal simultaneously, the resulting compensatory behaviour may be enough to explain "circling" and "head-butting" behaviours that have, until now, been thought of as existing independently.

Initially, we had wanted to create two Lego robots with the same goal of hitting a target on the other animal's body. However, technical challenges, along with my lack of experience, eventually resulted in the creation of 1 robot that would "attack" a target of a certain colour (green). The final robot consisted of 3 sets of wheels and corresponding motors situated at 120 degrees to one-another, and was able to move straight along the x and y axes, as well as rotate (and combinations thereof). It had 2 ultrasonic sensors placed at 180 degrees to one another on either side of its "head" and an iPhone strapped on the front providing live video feed. The "brain" of the robot was constructed using IQR, a large-scale neural simulator. It consisted of 4348 linear threshold neurons, 865 integrate and fire neurons, and 4 randomly spiking neurons. Sensor input was fed directly into populations of neurons which were arranged in a Winner Take All, and motor commands were issued depending on the current state. The behaviour of the robot consisted of moving toward the target item (a box with red and green sides). Once a certain threshold for "red" was surpassed (once the robot was close to the target), the behaviour would switch to circling to the left (i.e., moving around the target item to get to a different side). Since green was the target colour, the robot would accelerate toward green once it was detected at sufficient levels. The ultrasonic sensors produced lateral movement that served as collision avoidance, and in the absence of any sensor data, random motor commands were issued.

As a first approximation of targeting behaviour in cockroaches, the project was modestly successful. Future directions include making the sensory equipment of the robot more similar to that found in cockroaches (e.g., not relying on visual input), and making the neural architecture more sophisticated and biologically plausible. Additionally, the body plan of the robot might need to be modified so that behaviour emerging as a by-product of morphology could be observed.