Iterative and directed exploration of self-structuring embodied agents

Abstract

Roboticists increasingly look to biological systems for inspiration when designing and improving robotic systems. Simultaneously, building bio-inspired robotic systems can aid in the understanding of fundamental biological principles. In this context, the concept of embodied intelligence hypothesises that intelligent behaviours are the result of complex interactions between the brain, body (or morphology) and environment, rather than being driven purely by computational power in the brain. At the most basic level, embodied intelligence is driven by real-world physical interactions. By harnessing these interactions, unconventional `brainless' robotic systems have demonstrated complex behaviours driven purely by passive interactions. This thesis explores how complex behaviours emerge from interactions in two different low-level physical systems: falling paper and Bernoulli-balls. In falling paper systems, different paper shapes exhibit a range of behaviours when released into free fall. By altering morphological properties such as shape and weight, different behavioural modes can be triggered. In Bernoulli-ball systems, a ball is placed into a vertical airflow. If the morphological properties of the ball, for example size and density, and the environmental properties of the airflow, for example speed and width, are combined appropriately, the ball exhibits self-stable hovering within the airflow.

In Part I, I investigate falling paper systems. I introduce the novel V-shaped falling paper system. The relationship between morphology and system behaviours is explored and a data-driven modelling approach is developed to understand this. I explore the nature of behaviour transitions in the system. Certain behaviour transitions appear random, while others are more deterministic, and this variability is linked to morphology. Different methods are developed to represent this. I investigate generalised falling paper systems via the development of an automatic experimental platform capable of fabricating, dropping, observing and modelling hundreds of different paper shapes. Since falling paper systems are challenging to model using conventional methods, combining a data-driven approach with automatic experimentation is powerful.

In Part II, I investigate Bernoulli-ball systems. I explore the behaviour of a single Bernoulli-ball. A reduced-order model is derived to represent the main dynamics, and a minimalistic control policy is developed to modulate the ball hovering height by changing the airflow properties. I introduce the novel concept of a collective Bernoulli-ball: multiple hovering balls in a single airflow. This collective system exhibits a range of agent- and population-level behaviours, and these are investigated. The stability of, and relationship between, different behaviours is shown to be dependent on the balloon morphology and the environmental properties of the airflow.

In summary, the work in this thesis relates to the emergence of non-trivial behaviours from low-level embodied physical systems. The main contributions are the investigation of novel dynamics in these systems and the development of methods for understanding, representation and design. Ultimately, the work represents a small step toward the goal of creating artificial lifeforms with increasingly complex behaviours.