Minimally sparse but maximally versatile behavioural controllers - Mining for minimal controllers of intelligent locomotor behaviour

Research Areas: 

Animals like stick insects show smooth and fast transitions between behavioural states. So they can adapt motion sequences to the context and their behaviour to ontogenetic and phylogenetic changes.
The objective of this project is the identification of invariant features in the whole-body coordination of climbing and walking stick insects of different species, sex and age. From this we will try to derive a minimal set of behavioural controllers for different contexts, developmental stages and constraints.

Methods and Research Questions: 
  1. How to describe most efficiently multi-segmental coordination during locomotion as a sequence of discrete “motion states”?
  2. What are the invariant features in inter-limb coordination across stick insect species of different body shape, habitat or behavioural repertoire?

It is reasonable to assume that the basic principles of multi-legged locomotion can be identified and studied in each species of multi-legged walking animals. Therefore, comparing the mechanisms that underlie the control of locomotion in different species is likely to reveal invariant features. As yet, most investigations on insect walking have been limited to a small number of species. For example, the stick insect Carausius morosus has become an important model system for the study of autonomous, multi-legged locomotion in general. Here we compare multi-segmental coordination of stick insect species of different size (age, species) and body shape (species, sex) in order to identify invariant features in spite of morphological or ecological constraints. At the same time, this will help to identify particular adaptations to these constraints.

In a series of motion capture experiments, whole-body movements of stick insects are recorded that provide time-courses of more than thirty degrees of freedom (DoF) in the legs, thorax, neck and antennae. The continuous variables are then discretised into functionally meaningful bins (e.g., retraction, protraction, no movement). Because the joints of insects are actuated by antagonistic muscles that are activated by more or less discrete bursts of neural activity, discretisation or even binarisation of the measured variables is a valid reduction of data complexity. The outcome of discretisation essentially amounts to modelling the real, continuous, neural controller as a finite state machine with a finite set of discrete behavioural output states, each of which is characterised by a vector of movement states per DoF and/or activity states per muscle. Finally, the reduced data set is searched for a minimal set of discrete action states that is sufficient for generation of versatile motor behaviour.


The first data for the data set was collected with a step-climbing paradigm. The 3D-coordinates of 18 markers attached to the different segments of the animals were recorded with a frequency of 200 Hz. The time courses of all angles and angular velocities were calculated from the positions of markers attached to the animals (upper figure). The experiment was performed with four different step heights and repeated ten times for each animal and each step height. For a start, male Aretaon asperrimus were studied, expanding an existing data set on female Carausius morosus. These species differ strongly in body shape, regularity of locomotion, reproductive behavior and habitat.  At present, more than 500 trials for each one of these species were compiled for a Matlab data base. The lower figure shows a climbing sequence of Aretaon asperrimus as a set of superimposed stick figures.