Can music emerge from a swarm of robots each playing a single note and coordinating its behaviour with the others? We explore this idea by proposing a modular framework for the emergent generation of music, representing a novel intersection between robotics and artistic creation. We move beyond the works that link sound to robot movements or that allocate robots to musical roles. In our system, despite being limited to playing the atomic musical element, i.e., a single note, robots self-organise to play musical creations collectively. We illustrate the modular architecture of our framework by presenting three independent modules that run in parallel to enable the swarm to reach (i) temporal coordination so that robots play in synchrony, (ii) harmonic consensus so that notes are harmonically coherent, and (iii) beat distribution so that notes are distributed throughout time. We implement algorithms for the three modules building upon and extending existing swarm robotics solutions. Our bottom-up and modular approach also enables the use of cheap and accessible robots, hence fostering applicability, scalability, and robustness. Finally, combining the robot’s physical embodiment with the swarm’s plurality brings a unique dimension to the musical performance. We showcase our collaborative music creation framework with simulations and a real robot performance comprising 12 robots. This study shows the potential of combining music with swarm robotics to create musical complexity from simple robotic actions.

In this paper, we study the aggregation of a relatively high number of C. elegans roundworms resulting from the interaction between them and a set of external stimuli, namely oxygen and pheromones. Inspired by previous work, we delve into a hybrid modelling approach, simulating the aggregating behaviour of the worms by means of mathematical models, which we solve numerically, and by developing a phenomenological agent-based simulator. Our approaches capture the emergent aggregating behaviour of the worms, resulting in an interesting and coupled interplay between the macroscopic level, where probability distributions measure how the worms aggregate, and the microscopic level, where simple rules guide the agents to form clusters. Overall, our results suggest that there exists a strong correlation between the two approaches, indicating that they are able to capture the same phenomenon, albeit the mathematical model suffering from numerical limitations and the simulator requiring high computational resources. We then leverage our simulator in order to analyse the hypothetical behaviour of worms subject to both stimuli with different response levels, which results in a stronger degree of clustering than the single-stimulus simulation. We believe our framework can help shed light on the complex interactions between very large swarms of C. elegans , helping in predicting the emergent properties and reducing the amount of resources needed to run experiments with natural worms.

We propose a novel generic information-theoretic framework for characterizing the task difficulty in the Collective Perception paradigm. Our formalism builds on the notion of Empowerment - a task-independent, universal and generic utility function, which characterizes the level of perceivable control an embodied agent has over its environment. Series of simulations with an empowerment model of the collective perception scenario revealed a significant correlation between the levels of empowerment and the accuracy demonstrated by a set of standard collective decision-making strategies and a recent state-of-the-art neural network controller on nine benchmark patterns, used previously for assessing swarm performance. The results elucidate the key role of both the agent embodiment and the environmental pattern in characterising task difficulty, and justify the application of empowerment to analytically assess this role, which could help predict swarm performance and support the development of more efficient decision-making strategies.

Self-organised aggregation, the formation of large clusters of independent agents, is an important process in swarm robotics systems since it is the prerequisite for more complex collective behaviours. Previous work on self-organised aggregation focused on the study of the individual mechanisms required to allow a swarm to form a single aggregate. In this paper, we discuss an analytical model which looks at the possibility to use the concept of informed individuals to allow the swarm to distribute on different aggregation sites according to proportions of individuals at each site arbitrarily chosen by the designer. Informed individuals are opinionated agents that selectively prefer an aggregation site and avoid to rest on the non-preferred sites. We study environments with two aggregation sites, and consider two different scenarios: one in which the informed individuals are equally distributed in numbers between the two sites; and one in which informed individuals for one type of site are three times more numerous than those on the other site. Our objective is to find out whether and for what range of model parameters the swarm distributes between the two sites according to the relative distribution of informed agents among the two sites. The analysis of the model shows that the designer capability to exploit informed individuals to control how the swarm aggregates depends on the environmental conditions. For intermediate values of the site carrying capacity, a small minority of informed individuals is able to guide the dynamics as desired by the designer. We also show that the larger the site carrying capacity the larger the total proportion of informed individuals required to lead the swarm to the desired distribution of individuals between the two sites.

The work presented in this paper aims to address the problem of autonomous driving (especially along ill-defined roads) by using convolutional neural networks to predict the position and width of roads from camera input images. The networks are trained with supervised learning (i.e., back-propagation) using a dataset of annotated road images. We train two different network architectures for images corresponding to six colour models. They are tested “off-line” on a road detection task using image sequences not used in training. To benchmark our approach, we compare the performance of our networks with that of a different image processing method that relies on differences in colour distribution between the road and non-road areas of the camera input. Finally, we use a trained convolutional network to successfully navigate a Pioneer 3-AT robot on 5 distinct test paths. Results show that the network can safely guide the robot in this navigation task and that it is robust enough to deal with circumstances much different from those encountered during training.