The long term vision of the Autonomous Robot Evolution (ARE) project is to create an ecosystem of both virtual and physical robots with evolving brains and bodies. One of the major challenges for such a vision is the need to construct many unique individuals without prior knowledge of what designs evolution will produce. To this end, an autonomous robot fabrication system for evolutionary robotics, the Robot Fabricator , is introduced in this paper. Evolutionary algorithms can create robot designs without direct human interaction; the Robot Fabricator will extend this to create physical copies of these designs (phenotypes) without direct human interaction. The Robot Fabricator will receive genomes and produce populations of physical individuals that can then be evaluated, allowing this to form part of the evolutionary loop, so robotic evolution is not confined to simulation and the reality gap is minimised. In order to allow the production of robot bodies with the widest variety of shapes and functional parts, individuals will be produced through 3D printing, with prefabricated actuators and sensors autonomously attached in the positions determined by evolution. This paper presents details of the proposed physical system, including a proof-of-concept demonstrator, and discusses the importance of considering the physical manufacture for evolutionary robotics.

Swarms consist of many agents that interact according to a simple set of rules, giving rise to emergent global behaviours. In this paper, we consider swarms of mobile robots or drones. Swarms can be tolerant of faults that may occur for many reasons, such as resource exhaustion, component failure, or disruption from an external event. The loss of agents reduces the size of a swarm, and may create an irregular structure in the swarm topology. A swarm’s structure can also be irregular due to initial conditions, or the existence of an obstacle. These changes in the structure or size of a swarm do not stop it from functioning, but may adversely affect its efficiency or effectiveness. In this paper, we describe a self-healing mechanism to counter the effect of agent loss or structural irregularity. This method is based on the reduction of concave regions at swarm perimeter regions. Importantly, this method requires no expensive communication infrastructure, relying only on agent proximity information. We illustrate the application of our method to the problem of surrounding an oil slick, and show that void reduction is necessary for full and close containment, before concluding with a brief discussion of its potential uses in other domains.

In nature, group behaviours such as flocking as well as cross-species symbiotic partnerships are observed in vastly different forms and circumstances. We hypothesize that such strategies can arise in response to generic predator-prey pressures in a spatial environment with range-limited sensation and action. We evaluate whether these forms of coordination can emerge by independent multi-agent reinforcement learning in simple multiple-species ecosystems. In contrast to prior work, we avoid hand-crafted shaping rewards, specific actions, or dynamics that would directly encourage coordination across agents. Instead we test whether coordination emerges as a consequence of adaptation without encouraging these specific forms of coordination, which only has indirect benefit. Our simulated ecosystems consist of a generic food chain involving three trophic levels: apex predator, mid-level predator, and prey. We conduct experiments on two different platforms, a 3D physics engine with tens of agents as well as in a 2D grid world with up to thousands. The results clearly confirm our hypothesis and show substantial coordination both within and across species. To obtain these results, we leverage and adapt recent advances in deep reinforcement learning within an ecosystem training protocol featuring homogeneous groups of independent agents from different species (sets of policies), acting in many different random combinations in parallel habitats. The policies utilize neural network architectures that are invariant to agent individuality but not type (species) and that generalize across varying numbers of observed other agents. While the emergence of complexity in artificial ecosystems have long been studied in the artificial life community, the focus has been more on individual complexity and genetic algorithms or explicit modelling, and less on group complexity and reinforcement learning emphasized in this article. Unlike what the name and intuition suggests, reinforcement learning adapts over evolutionary history rather than a life-time and is here addressing the sequential optimization of fitness that is usually approached by genetic algorithms in the artificial life community. We utilize a shift from procedures to objectives, allowing us to bring new powerful machinery to bare, and we see emergence of complex behaviour from a sequence of simple optimization problems.

Referential communication is a ”representation-hungry” behavior, and the bee waggle dance is a classical example of referential communication in nature. We used an evolutionary robotics approach to create a simulation model of a minimalist example of this situation. Two structurally identical agents engage in embodied interaction such that one of them can find a distant target in 2D space that only the other could perceive. This is a challenging task: during their interaction the agents must disambiguate translational and communicative movements, allocate distinct behavioral roles (sender versus receiver), and switch behaviors from communicative to target seeking behavior. We found an evolutionary convention with compositionality akin to the waggle dance, correlating duration and angle of interaction with distance and angle to target, respectively. We propose that this behavior is more appropriately described as interactive mindshaping, rather than as the transfer of informational content.

Human judgement is better described as a heuristic process rather than maximisation of a utility function. Heuristics are high cognitive processes of decision-making whose rapid and effortless implementation is useful to confront risk scenarios that compromise the viability of an individual. They can be defined in an enactive frame as self-sustained and self-generated habits of abductive behaviour selection in sensorimotor agents influenced by the individual history of sensorimotor contingencies and the environment. In this work, we analyse the emergence of patterns of behaviour and its necessary ecological conditions when performed decisions are related to energy intake and energy expenditure. Agent’s sensors and intentions are coupled to a variation of the iterant deformable sensorimotor medium (IDSM) (Egbert and Barandiaran, 2014). This model explains transparently the generation of sensorimotor habits in simulated robots through the influence of registers of previously executed behaviours reinforced by repetition. We create a decision-making frame based on intentions as probabilities of specific actions which constitute the motor component of sensorimotor states on IDSM. In this model is seen that specific behaviour correlations with the lifespan of agents depend on the availability of energetic sources on the environment. Inheritance of the medium is introduced in agents with a small improvement on the lifespan of agents.