Insects are really cool
Last updated: Feb 1, 2022
It is easy to find amazing examples of insect behaviour. The process by which ants build nests is beautiful. Everything from hatching eggs to the incredible food gathering dynamics is awe inspiring. What drives these very complex behaviours in such energetically efficient systems?
Control mechanisms in insects
Control mechanisms behind insect behaviour are surprisingly complex. For instance, insects can exhibit Pavlovian traits, altering their behaviour depending on previous stimulus. Even more surprisingly, insects can track some environmental variables (for example, bees and crabs seem to posses “home vectors” to track the location of their nests). There is no single cognitive process behind these behaviours, instead, it appears cognition in insects is very complex (Cognition in insects, 2012). Furthermore, research has shown that insects can disambiguate the effect of their own movement with respect to the environment, and represent visual space for orientation and navigation (What insects can tell us about the origins of consciousness, 2015).
Despite the complexity of their control mechanisms, it appears to me that insect behaviour is not built on a (primarily) learning-based platform. Perhaps much of the control mechanisms are instead “hardcoded” into the hardware platform. Moreover, it appears that raw “computational power” is not what enables complex behaviour: flight and host search are possible from the thousands of neurons (The smallest insects evolve anucleate neurons, 2011).
There is a lot of evidence that the heavy-lifting in insect behaviour optimization is done through evolution of hardware platforms, which introduce biases so that “learning” (if any) is efficient. Intuitively this makes sense: learning is extremely expensive relative to the lifespan of insects.
Some questions for which I would love to find the answer:
Is the hardware platform (i.e. body) what enables such (relatively) simple control structure? How is it possible that a system with so many sensors can be controlled with relatively few computation units? Assume we have 100 units of energy for compute. How much energy should we spend on designing a good hardware platform and how much on designing its control law?
Can such a process (insect evolution) be replicated by humanity at a lower energetic cost? Can we steer this process into something that is useful for humanity (e.g. highway-making ants)?
Co-evolution of body and mind has been historically explored (Karl Sims’ 1994 work remains a personal favorite). This continues to be explored, resulting in impressive works (e.g. Embodied Intelligence via Learning and Evolution, 2021, Regenerating Soft Robots through Neural Cellular Automata, 2021). Yet, this area seems to be under-explored when compared to the amount of work on algorithms for optimization of Artificial Neural Networks (in Reinforcement Learning-like settings). What is the best way to co-optimize robot control and robot design?