Central pattern generators (CPGs) produce rhythmic activation patterns through neural circuits. These neural circuits are abundant in organisms and are ultimately responsible for neural activity that produces behaviors such as walking, breathing, flying, swimming which makes them interesting to apply to robots. One mystery is how natural central pattern generators contribute to the emergence of Dynamically Stable Locomotion and how these central pattern generators were evolved.
Here are some topics that can be explored within this theme:
- Evolving neural network CPG structures. In robotics, CPGs are usually either abstracted to differential equations or created using neural networks. The goal of this project is to investigate how to best implement and optimize a controller that has inherent rhythmic activation patterns, and how this type controller can be effectively used to create locomotive patterns for simulated robots. We recommend using a small world neural network as it innately creates patterned activity. The weights and topology of this network can be evolved using an evolutionary algorithm. How well the small world network performs could be compared to another CPG approach.
- Gait flexibility in bio-inspired evolving virtual robots. The aim for this option is to investigate transitions in gait pattern such as between walking and trotting. The student will work on classifying these gaits, testing what kinds of evolutionary pressure lead more often to gait transitions, and investigate how this behaviour is reflected in properties of the CPG. This will be done with a simulated robot using a newly developed robot controller based on highly non-linear neuron models (see COROBOREES project).
- Morphological dependence for synchronization. Rhythmic entrainment is a rare but cognitively significant ability in the animal kingdom. Current work on the COROBOREES project has succeeded in real-time synchronization of a quadruped to external stimulus (see video above). Questions remain regarding how this depends on body morphology. One approach to this question could be to interface the controller scheme with a modular morphology, where body parts are easily added or removed.
We imagine these projects being carried out mainly using simulations within the Unity framework, but we are also open to other ideas. It could also be a possibility to test out the developed methods on a hardware robotic platform.
