FocusCOLLECTIVE BEHAVIOR

Reflections on the future of swarm robotics

See allHide authors and affiliations

Science Robotics  09 Dec 2020:
Vol. 5, Issue 49, eabe4385
DOI: 10.1126/scirobotics.abe4385

Abstract

Swarm robotics will tackle real-world applications by leveraging automatic design, heterogeneity, and hierarchical self-organization.

The news is full of stories about robots conquering new grounds—a new application, a new operation environment, or a new level of performance. With rare exceptions, these stories are mostly about one or a few robots working together. In fact, the deployment of large groups of robots, or robot swarms, that coordinate and cooperatively solve a problem or perform a task, remains a challenge.

For the past two decades, overcoming this challenge has been the goal of swarm robotics, a research field that takes inspiration from natural self-organizing systems such as social insects, fish schools, or bird flocks and attempts to recreate the emergence of collective behavior from simple local interaction rules. Its goal is to make multi-robot systems more robust, fault-tolerant, and flexible than single robots.

The swarm robotics field finds its roots in a few seminal works published in the 1990s (1, 2), but it started to grow substantially only after the year 2000, moving from a small domain featuring studies with a clear biological inspiration to a mature research field involving many laboratories and researchers worldwide.

Nowadays, robot swarms have been demonstrated in lab settings, frequently employing a small number of robots with compact size (3, 4). Although technological innovations are pushing boundaries toward ever smaller scales (5, 6) and larger numbers (7, 8), the path toward real-word applications is still long and arduous. Future robot swarms will exist and operate at many different scales: physical scale—from micro/nano robots to large ground, aerial, and aquatic robots; group scale—from a few dozen to millions of individuals composing the swarm; temporal scale—from swarms that display fast intervention and quick adaptation in a rapidly evolving environment to robots that continuously operate on months-long missions (e.g., on a distant planet); and spatial scale—from deployments in small confined spaces to missions across expansive unbounded fields. To design and manage such a wide range of possible systems, the key challenge will be to define a rigorous engineering methodology to program the individual robots so that the swarm as a whole acts as desired. This is no easy task, because the characteristics of swarms at different scales might require radically different approaches.

To grasp the variety of possible systems and applications, we suggest a few general design principles that remain valid across scales. First, the increasing complexity of swarm systems is such that their design cannot be accomplished solely by traditional approaches. The more robot swarms will be confronted with uncertain/unpredictable environments and will rely on intricate patterns of interactions, the more automated design methodologies will be necessary to obtain desired behaviors, because they can be employed to generate individual rules that are evaluated for their effects on swarm performance (9). Machine learning with data-driven approaches becomes relevant whenever model-based solutions are too demanding, for instance, when it is difficult to provide a precise model of robot-environment interactions (e.g., due to complex physical interactions among micro/nano robots or unpredictable environmental dynamics as caused by underwater currents). To address self-organized control, the learned control architectures need to suitably integrate robot perceptions with information asynchronously received from (possibly hundreds of) peers and with memory of past states. Hybrid systems mixing model-free and model-based approaches will likely provide additional power, learning from data those aspects that are peculiar to the problem at hand (e.g., the response of a noisy communication medium characterizing the target application). Overall, automated methods hold the potential to liberate the designer from tedious trial and error or parameter tuning and to better deal with the specific swarm scales required by the task, resulting in a general engineering methodology that can be suitably applied across different application domains.

Second, in order for robot swarms to perform ever more complex tasks, they will likely need to be heterogeneous, both in hardware features and in roles that individual robots can play within the swarm (4). Physical heterogeneity can equip the robot swarm for tasks that require different hardware (e.g., fast moving drones for collaborative monitoring that cooperate with slow ground robots capable of modifying the environment). Behavioral heterogeneity can enable specialization (possibly via local learning mechanisms) and can underlie phase transitions that make the emergent collective behavior more flexible, adaptive to new conditions, and resilient to external perturbations. By abandoning the traditional homogeneity of hardware and control, flexibility and autonomy can be improved for small and large swarms, and operations can be expanded to address a wider spectrum of spatial and temporal scales. The additional dimensions introduced by heterogeneity (i.e., number of different roles and their proportion within the swarm) entail increased complexity in the design and further motivate the exploitation of automated techniques.

Last, robot swarms should include mechanisms to allow hierarchical forms of control beyond traditional pure self-organization. Whereas the latter features a flat organizational structure, a hierarchical control approach assigns a few individuals larger responsibilities, flexibly adapting the hierarchy to the task execution demands (10). In this way, it could be possible to more efficiently address task allocation, creation of task-oriented teams within the swarm, coordination of specific activities, or interaction with human users. Such hierarchies should, however, not be imposed from the outside. Rather, they should themselves be the result of self-organizing processes where some robots in the swarm may take leading roles depending on their specific characteristics, on the task being performed, and on the environmental conditions in which they are placed. In this respect, leveraging heterogeneity for robot swarm control—as advocated above—becomes even more consequential.

An (automated) engineering methodology for heterogeneous, hierarchically self-organized robot swarms is still missing but is crucially needed to successfully transition from laboratories to real-world applications (see Table 1). We believe that the first deployments of these unconventional robot swarms will be in precision agriculture, infrastructure inspection and maintenance, and non-combat military applications. Precision agriculture features expansive fields where different types of robots can operate in parallel and can collaborate in heterogeneous teams to perform important actions such as weeding, fertilizing, or harvesting. Infrastructure inspection can benefit from the parallel operation carried out by large numbers of small robots exploring different areas and collectively identifying the presence of faults or damage. Non-combat defense operations will benefit from swarms that—by exploiting hierarchical self-organization—are difficult to shut down because they can flexibly re-organize and self-repair. Later in the next decade, aquatic surface and underwater robots will be mature enough to be deployed at sea, carrying out inspection, ecological monitoring, deep-sea exploration, and sustainable fishing. Robot swarms will then likely enter our cities, taking over service and entertainment tasks in a coordinated and self-organized way, while interacting with human workers and citizens. The use of robot swarms in space missions could come next and could play a key role in large-scale exploration of the Moon and Mars by 2040.

Table 1 Timeline of swarm robotics research.

Milestones reached by swarm robotics research in the past and projections of future developments. For each milestone, we indicate whether it entails innovation in software (SW), hardware (HW), or both.

View this table:

Another possible future direction takes advantage of miniaturization and unconventional materials, possibly exploiting biohybrid solutions (11). Soft-bodied robots enable alternative forms of interaction and physical self-organization that can be exploited in an innovative way. For instance, swarms of soft-bodied, biodegradable, millimeter-scale robots could be deployed in agricultural fields, coordinating to attack pests or harmful soil nematodes without damaging the crops, or in the sea to harvest microplastics by self-assembling into macroscopic structures to be later collected by larger water-cleaning robots. The last—and more challenging—frontier would be medical applications, requiring extreme miniaturization and biocompatibility. Considering the current advancements (5, 6), nanoscale robot swarms could be ready for testing by the mid-21st century. The successful deployment of robot swarms in these and other applications could pave the way for self-organized, bioinspired swarm robotics to become a mainstay of engineering in the second half of the century, possibly establishing itself as the standard way of designing complex robotic systems.

REFERENCES

View Abstract

Stay Connected to Science Robotics

Navigate This Article