Designing collective behaviors for robot swarms is a difficult endeavor due to their fully distributed, highly redundant, and ever-changing nature. To overcome the challenge, a few approaches have been proposed, which can be classified as manual, semi-automatic, or automatic design. This paper is intended to be the manifesto of the automatic off-line design for robot swarms. We define the off-line design problem and illustrate it via a possible practical realization, highlight the core research questions, raise a number of issues regarding the existing literature that is relevant to the automatic off-line design, and provide guidelines that we deem necessary for a healthy development of the domain and for ensuring its relevance to potential real-world applications.

The reality gap-the discrepancy between reality and simulationis a critical issue in the off-line automatic design of control software for robot swarms, as well as for single robots. It is understood that the reality gap manifests itself as a drop in performance: when control software generated in simulation is ported to physical robots, the performance observed is often disappointing compared with the one obtained in simulation. In this paper, we investigate whether, to observe the effects of the reality gap, it is necessary to assume that the control software is designed in a context that is simpler than the one in which it is evaluated. In a first experiment, we show that a performance drop may be observed also in an artificial, simulation-only reality gap: control software is generated on the basis of a simulation model and assessed on a second one. We will call this second model a pseudo-reality. We selected the simulation model to be used as a pseudo-reality by trial and error, so as to qualitatively replicate previously published observations made in experiments with physical robots. The results show that a performance drop occurs even if we can exclude that pseudo-reality is more complex than the simulation model used for the design. In a second experiment, we eliminate the trial-and-error selection of the first experiment by evaluating control software across multiple pseudo-realities, which are sampled around the original simulation model used for the design. The results of the second experiment confirm those of the first one and show that they do not depend on the specific pseudo-reality we previously selected by trial and error. Moreover, they suggest that one could use multiple pseudo-realities to evaluate automatic design methods and, from this simulation-only evaluation, infer their robustness to the reality gap. The experiments were conceived by the two authors and performed by AL. The article was drafted by AL and revised by the two authors. The research was directed by MB.

Neuro-evolution is an appealing approach to generating collective behaviors for robot swarms. In its typical application, known as off-line automatic design, the neural networks controlling the robots are optimized in simulation. It is understood that the so-called reality gap, the unavoidable differences between simulation and reality, typically causes neural network to be less effective on real robots than what is predicted by simulation. In this paper, we present an empirical study on the extent to which the reality gap impacts the most popular and advanced neuro-evolutionary methods for the off-line design of robot swarms. The results show that the neural networks produced by the methods under analysis performed well in simulation, but not in real-robot experiments. Further, the ranking that could be observed in simulation between the methods eventually disappeared. We find compelling evidence that real-robot experiments are needed to reliably assess the performance of neuro-evolutionary methods and that the robustness to the reality gap is the main issue to be addressed to advance the application of neuro-evolution to robot swarms.

Previous research has shown that automatically combining low-level behaviors into a probabilistic finite state machine produces control software that crosses the reality gap satisfactorily. In this paper, we explore the possibility of adopting behavior trees as an architecture for the control software of robot swarms. We introduce Maple: an automatic design method that combines preexisting modules into behavior trees. To highlight the potential of this control architecture, we present robot experiments in which we compare Maple with Chocolate and EvoStick on two missions: foraging and aggregation. Chocolate and EvoStick are two previously published automatic design methods. Chocolate is a modular method that generates probabilistic finite state machines and EvoStick is a traditional evolutionary robotics method. The results of the experiments indicate that behavior trees are a viable and promising architecture to automatically generate control software for robot swarms.JK and AL contributed equally to the research and should be considered co-first authors. Behavior trees were originally brought to the attention of the authors by DB. The proposed method was conceived by the four authors. It was implemented and tested by JK and AL. The initial draft of the manuscript was written by JK and AL and then revised by DB and MB. The research was directed by MB.The final authenticated version is available online at https://doi.

Optimisation-based design is an effective and promising approach to realising collective behaviours for robot swarms. Unfortunately, the domain literature remains often vague on the exact role played by the human designer, if any. It is our contention that two cases should be disentangled: semi-automatic design, in which a human designer operates and steers an optimisation process (e.g., by fine-tuning the parameters of the optimisation algorithm); and (fully) automatic design, in which the optimisation process does not involve, need, or allow any human intervention. In the paper, we briefly review the relevant literature, we illustrate the hypotheses, the characteristics, and the core challenges of semi-automatic and automatic design, and we sketch the context in which they could be ideally applied.