Summary

Swarms are on the cusp of being a practical reality. They offer a new way of delivering functionality, with many parts coming together to form a whole. But challenges remain: designing swarm behaviours – which are not a property of the component sub-units but only emerge from the entire swarm system – is a difficult task. For now, readers would be right to be sceptical of bold claims: what is possible in a lab far outstrips swarms' maturity in uncontrolled environments. And even if/when swarms do become realistically useable in operational environments, the cost and complexity of procuring and operating them may take away from their benefits.

Nevertheless, swarming technology does offer some compelling advantages – speed, volume, resilience – and it is an area of active and growing research interest both in the UK and for our partners and competitors. Developments in other industries mean that the components needed for swarms are likely to be available at suitably low prices, and assembling them into at least a basic swarm will not be a great technical challenge. Guaranteeing that the swarm works as intended (that point about designing behaviours) might take a little longer, particularly in complex environments.

Swarm research and development is ongoing; thinking about how swarms might practically be used is less well-advanced. Application-specific thought on how swarms may be controlled, and how they may work alongside humans and other autonomous systems will aid potential users in preparing the way for adoption.

Introduction

The swarm. A paradoxical idea of huge collective scale and small individual size. A staple of modern science fiction. And now a technological reality, promising the ability to marshal vast numbers of bots in service of a single goal.

So it would seem, at least, if we believe the breathless popular science articles that pop up with reasonable regularity.Hollings, Alex, 'Watch: Raytheon shows how drone swarms and lasers will soon make warfare a sci-fi horror show', SOFREP, February 2019, https://sofrep.com/news/watch-raytheon-shows-how-drone-swarms-and-lasers-will-soon-make-warfare-a-sci-fi-horror-show/ But what’s the reality underneath the hype? And what, realistically, are the benefits that swarms might bring?

What is a swarm?

Before we can answer those questions, though, we need to know what we’re actually talking about – what is a swarm?

Perhaps it's easier to start with what a swarm isn't: it isn't just a big group of things (we’ll use the technical term nodes). There's an important conceptual distinction: when we think of a group, the primary 'building blocks' are the members of the group – they separately provide capability that adds up to a total effect; when we think of a swarm, the swarm itself is the primary building block, with a capability that is distributed across the swarm.

A swarm is a collective, built up from nodes that each have some discrete amount of functional capability. The nodes work together following a defined set of rules, collectively producing a whole that has qualities and a total level of capability that are greater than the sum of its parts, a property called emergence.Beni, Gerardo, ‘From Swarm Intelligence to Swarm Robotics’, Proc. of Workshop on Swarm Robotics, 8th Int'l Conf. on Simulation of Adaptive Behavior (SAB '04), 13-17 July 2004, https://doi.org/10.1007/978-3-540-30552-1_1

The nodes that make up a swarm are interchangeable, although not necessarily completely identical. They may temporarily take on different roles within the swarm; they may have some degree of specialism; or a swarm might be made up of several distinct sub-populations – but a significant degree of interchangeability must exist for the group to be considered a true swarm.

Swarms exist in nature as a result of many individual animals following a shared set of rules for how they behave. In many cases they offer the group as a whole some sort of advantage: making it more difficult to predators to catch them, bringing individuals together to mate, or making migration less energy-intensive. But although the whole group benefits, the natural swarms are not guided by any central intelligence: their behaviour is determined by their internal rules and their boundary conditions – in other words, their environment. In contrast, technological swarms do need some way for a user to direct or control them if they're to be practically useful. That control may be continuous or momentary (for example loading a set of instructions that are then followed autonomously), and is exerted over the swarm as a whole, not individual nodes: swarm nodes are autonomous. Control can be achieved in several ways:Verbruggen, Maaike, ‘The question of swarms control: challenges to ensuring human control over military swarms’, Non-Proliferation and Disarmament Paper no. 65, December 2019, https://www.nonproliferation.eu/wp-content/uploads/2019/12/EUNPDC_no-65_031219.pdf

• the algorithms that each node uses to decide on its actions can be switched – for example switching from a loitering mode to a travelling mode;
• the parameters or objectives those algorithms use can be changed – for example setting the maximum speed for a node or a target set of co-ordinates;
• or the swarm's environment can be controlled, forcing it to respond in the desired way – for example placing a barrier in the swarm's path.

Why swarms?

Before getting any deeper into the details of swarms, there’s an important question we need to answer: why should anyone be interested?

Many of the benefits of swarms come from the fact that they are made up of a large number of small nodes:

• Swarms make quick work of big jobs. The many individual nodes in a swarm mean a task can be divided into many little chunks, with each node working on its own chunk at the same time as all the others are working on theirs. In many cases, that means a swarm can do a job faster than a single, more sophisticated entity. Say you want to take high-resolution aerial photographs of an area: a 100-strong swarm would in principle be able to finish quicker than a single larger aircraft.

• Swarms provide volume. Swarms multiply the rate at which work can be done – every node can work in parallel with the others. Swarms also offer a more straightforward sort of volume: weight of numbers.
• Swarms are (sometimes) cheap. The relationship between cost and capability is non-linear, so it can often be less expensive to buy ten units of a basic system than to buy one unit of a system with ten times the capability. If you can combine those ten units into a swarm that has the same overall effect, then the swarm can provide the same amount of capability at lower cost. But the economics may not always work out.

A lot of these benefits can be realised with any large group, but the addition of a swarm's emergent properties, the interchangeability of its nodes, and its autonomy multiplies and extends them:

• Swarms are resilient. If one node stops working, it has only a small impact on the outcome because others can step in to replace it, self-organising to re-allocate tasks.
• Swarms are flexible. A swarm’s size and shape are not fixed, but can be reconfigured to meet changing needs. This can be dynamic, the swarm adapting responsively to the context, or it can be task-dependent, with a single swarm carrying out a multi-phase task and reconfiguring itself between phases.

Swarms, especially technological ones, are perhaps most readily imagined as airborne things. But in fact, swarms can operate in any physical domain (on the ground, on or under the sea, or in space), and there are analagous ideas in the digital world (like distributed denial of service attacks).

Swarm applications

The inherent benefits of swarms make them particularly well-suited to applications in information-gathering and transport uses. There are also headline-grabbing military applications.

Information-gathering

A swarm’s distributed capability means it can quickly cover a large area at close range: an airborne swarm could collect close-up aerial imagery and other sensor data to rapidly build a multi-layered, high-resolution map of an area – or it could fly around and through a large structure like a bridge to inspect it for defects. It also means a swarm can collect measurements at the same time from different places, which can be useful for the development of accurate digital models or to reduce the likelihood of missing a moving target of interest in the area being surveyed.

The interchangeability of a swarm’s nodes means that a swarm can have an effective endurance far longer than that of any individual node. By continuously retiring and replacing nodes for refuelling or recharging, a swarm might be able to provide extended-duration coverage of an area.

To be effective in an information-gathering application, a swarm will need to be integrated with other parts of a wider system. It will need to get to and from the location of interest, it will need to be controlled, and the data it collects will need to be retrieved:

• Movement. A swarm could travel from a fixed home base to a location of interest under its own power, but it is more likely that it will be transported at least some of the distance in some other vehicle. The way in which it is transported, deployed, and recovered will all need consideration.
• Control. It may be advantageous for a swarm to complete its information-gathering task without the need for any control link, or the user may want to have the ability to monitor it and adjust mid-task: either way, it’s important to start with a good understanding of the requirement for control over the swarm.
• Data. The time-sensitivity of the data and the availability of a suitable communications link will determine whether data is sent back to a home base on the fly or if it is simply downloaded at the end of the task. In this latter case, it is important to consider the way data is stored within the swarm, both to avoid data loss should a node fail and to simplify the data download process.

Transport

A swarm can, collectively, carry a large load, can deliver it to one place or many, and can do so all at once or over a longer period of time. A fire-fighting swarm could draw on water from a single source to tackle fires in multiple places, or could support a much more continuous flow of water to one fire than a single, larger firefighting platform. A seed-planting swarm can take a supply of seeds and quickly distribute them across a field – or a similar system could deposit remote sensor packages over a wide area (or equally, the nodes themselves could become the sensors).

Swarms for transport will require similar consideration of the appropriate approach to control over them. The introduction of a physical interface with some wider logistic system – the point at which the thing being transported is transferred to the swarm – will requires some additional design effort. The large number of nodes in a swarm means that any friction at that interface, like a need for human operator intervention, will be amplified.

Military

The most headline-grabbing military application of swarms is in an offensive role. The combination of the large number of nodes and the responsiveness of swarms makes them seem a useful tool for conducting attacks against strong defences. In reality, nodes will be constrained to small sizes by cost concerns, limiting the firepower they can individually carry, and the complexity of a combat environment will present a major challenge for autonomous operation.

How to build a swarm

So you've got an application that sounds like a good fit for a swarm: does that mean it's time to start building? Well, no – first you need to double-check if a swarm is actually the right solution.

Is a swarm right for the job?

Swarms build up an overall capability from many small chunks – but how small? Before you start building your swarm, you have to understand what the right size for those chunks is. In some cases, it might not be possible to divide up the capability to a sufficient extent to make a swarm approach make sense – or it might not be possible to divide it up at all.

There are other reasons a swarm-based solution might not make sense, and these too need consideration. Large numbers of nodes mean that per-node costs can quickly mount up:

• if the total cost turns out to be higher than for a single platform that can meet the need then a swarm will not be cost-effective;
• if the expected rate at which nodes may fail or be lost is high, the cost to replace them might make operating a swarm too expensive;
• if the cost of powering or of refuelling the nodes is too high;
• nodes need maintenance – the cost or effort needed for that may be too great.

Even if the cost implications of a large number of nodes do work out, the extra development effort involved in producing robust, controllable true swarming behaviour may mean that other ways of controlling large numbers of nodes are more attractive – like pre-programming them or using some form of remote control, resulting in a group that is not a true swarm.

Design choices

Once you’ve confirmed a swarm approach is right for your particular problem, you need to figure out a command and control architecture. You could pre-program your swarm to carry out a task; have every node connecting back to a central control hub; use a defined hierarchy within the swarm; use a dynamic mesh network; or employ some combination of all of these. The choice of swarm architecture will depend on the communications channels you’ll have available to you, practical considerations around the communications abilities of each swarm member, and on how flexible or responsive you want your swarm to be. The need for some form of fail-safe behaviour, the degree of autonomy desired and whether that's fixed or variable, and the requirement for real-time data and control links will also influence the choice.

Some of the key advantages and disadvantages of these different architectures are summarised in the table below:

The choice of control architecture will also depend on the level of continuous control needed over the swarm’s actions. There is a conceptual challenge here,too: swarms are not controlled by passing commands directly to swarm nodes, but by setting parameters within which the swarm as a whole operates. If direct control over some action the swarm might take is needed, you need to decide whether you want the swarm to seek permission for each instance of that action it 'wants' to undertake; pre-authorise action subject to certain criteria being met; or take an approach somewhere in between. This challenge is shared with other autonomous systems, but is particularly acute for swarms given the defining lack of direct control over individuals.

Energy

Another important early design consideration for your swarm is how its energy needs will be managed. Each node will require energy: to move, to run processors and communications equipment, and to power sensors. The size, weight, and cost of each node will need to be balanced against the length of time it can power itself for. Depending on the application, nodes may be able to return to a base and refuel or recharge mid-mission, or they may need to carry sufficient energy to complete it. It may also be possible to effectively share energy within the swarm, either by transferring energy between nodes or by dynamically assigning energy-intensive tasks to nodes with the most energy remaining.Melhuish, C. and Kubo, M., 'Collective Energy Distribution: Maintaining the Energy Balance in Distributed Autonomous Robots using Trophallaxis', Proceedings of 7th International Symposium on Distributed Autonomous Robotic Systems, June 2004, https://link.springer.com/chapter/10.1007/978-4-431-35873-2_27

Simulation

One way to generate evidence on whether a swarm-based solution is appropriate to your problem, and if so what architecture might work best, is to simulate it. Simulation can also help with understanding how swarms can be integrated with other elements of a wider system – or how swarms operated by other people (including people with hostile intent) might affect existing systems.

The technology to simulate swarms is actually more mature than real-life swarming technology, because many of the barriers to development progress for swarms are related to practicalities: things like measurement errors, logistics, or the large up-front cost of building an experimental swarm. Swarm simulations have seen significant academic and industry effort, with a range of proprietary and open-source tools now available.Soria, E., Schiano, F., and Floreano,D., 'SwarmLab: a Matlab Drone Swarm Simulator', 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020, pp. 8005-8011, doi: 10.1109/IROS45743.2020.9340854, https://ieeexplore.ieee.org/document/9340854 In general, these have been integrated into current military simulation tools only in a limited, experimental way: this is a due to a combination of limited demand and technical challenge (although individual node behaviours can be simulated easily, achieving a simulation of emergent behaviour that matches what happens in the real world is difficult)

Practicalities

After you’ve got your swarm architecture sorted, you can move on to the details of designing and building the swarm. There are still challenges here: you need to prioritise keeping the unit cost of swarm members down if you’re to make a swarm approach cost-effective. Fortunately, some current technology trends can help you out there...

Underlying tech

Each swarm node is actually a pretty complex system in its own right. It has to fulfil several important functions, and do it all at minimal cost: it will need some sort of navigational sensing, so that it knows where it is relative to its environment and relative to the other swarm members; a communications package to allow it to talk to its peers and/or connect back to a command hub; processing capability to run the control algorithms and deal with all the data from the sensors; some means of motion control; a power source to keep everything running; and of course some sort of payload to provide the basic functionality you want the swarm for in the first place – perhaps a camera to take photos, a seed dispenser to sow plants, or a firefighting payload.

Navigational sensing

Nodes will in almost all cases need an understanding of their position and surroundings for navigation and collision avoidance. The accuracy provided by low-cost satellite positioning or inertial navigation modules isn't good enough for a swarm, but these can be combined with a wide range of readily-available sensors to improve accuracy: air pressure sensors; ultrasonic/laser distance measurement for altitude sensing; and magnetic field sensors to determine orientation, among others. There are many sensors of these sorts readily and cheaply available, developed for use in consumer devices like smartphones or retail drones. Even greater accuracy is promised by specialist 3D-scanning sensors being developed for applications like augmented reality devices or self-driving cars, and by the application of real-time processing of feeds from cameras to generate 3D models of the node's environment.

Communication within the swarm can also aid in refining and extending each node's awareness of its surroundings, by sharing sensor data and using collaborative approaches to arrive at a shared understanding, reducing the effect of individual measurement errors.

Communications

Data-sharing within the swarm is one reason that every node needs some sort of communications capability: some or all nodes may also need the ability to communicate with a command hub. There are a variety of already-developed communications standards and systems to allow in-swarm networking, including using dynamic mesh approaches. These have been, and continue to be, developed for applications in consumer electronics such as smart home devices and industrial 'internet of things' applications, both large and growing markets that are likely to drive further development and cost reduction.

The longer-range communications requirements for connections back to a command hub are a shared with many already-existing systems. The major difference is the even greater imperative to focus on cost.

Processing

Depending on the complexity of the sensor suite and the control algorithm, substantial processing ability may be required. And this has to be delivered with constraints on size, weight, electrical power consumption, and cost. The proliferation of smartphones has established a large demand for cheap, low-power, high-performing processors, and similar problems are faced in many edge computing systems: demand for these is likely to drive a continued expansion in the availability of suitable components. Looking a little further ahead, trends for low-cost application-specific chips and novel low-power computing technologies will also enhance the practically-achievable processing ability for swarm nodes.

Motion control

Any swarm needs the ability to move in the domain in which it operates: the specifics of how that is achieved will of course depend on the domain. The air is perhaps the domain most closely associated with swarms, and here the significant market in commercial and retail drones is driving increased availability of low-cost flight control systems and motors.

Power

As with any vehicle, swarm nodes need a power source. It needs to be cheap, of course, but also easily (and preferably autonomously) replenishable. Depending on which domain the swarm is to operate in, there will be other key requirements. In most cases, the power source will be the limiting factor on the length of mission the swarm can undertake: it's therefore particularly important to think carefully about whether sufficient performance can be achieved for an acceptable level of cost. In many cases, swarm nodes will be powered by batteries, and can therefore expect to benefit from trends for increasing battery energy density driven by consumer devices and electric vehicles.

Payload

Last, but by no means least, swarm nodes will need some sort of payload to deliver the functionality that is the reason for having the swarm in the first place. While there is a wide range of possibilities for what this might be, information-gathering is perhaps a prominent example. Here, nodes can benefit from the same trends that are set to increase the capability of the sensor suites available (described above) as well as trends supporting the development of more compact, less expensive sensors for small satellites, industrial automation, and the broad 'internet of things'.

Integration

After picking out the components and building your nodes, you still have to face the real technical challenge: making everything work together. That’s the final, critical piece of the puzzle – and while everything so far has been relatively straightforward, here we need to turn to the world of research and development.

The state of the art

In principle, and depending on the task, it shouldn’t be too difficult to get a swarm up and running. It is a complex task, requiring some in-depth system-of-systems thinking, but it seems like it ought to be doable. Unfortunately, while we’ve got quite good at designing control systems for things that work by themselves, those systems have a tendency to get unstable when you bundle them together in big groups – a problem compounded by the latency and error that are unavoidably present in real-world sensing and communication.Liu, Yang, Passino, K.M., Polycarpou, M., ‘Stability analysis of one-dimensional asynchronous swarms’, IEEE Transactions on Automatic Control, vol. 48, no. 10, pp. 1848-1854, October 2003, https://www.doi.org/10.1109/TAC.2003.817942

So while we’ve just about got the hang of building the algorithms that can make big swarms work in simulations, we’re only just starting to see those translated into real, physical swarms. Some of the headline-grabbing claims of huge swarms don’t really stack up if you look at the detail: they might have involved large numbers of nodes, but they often involve some degree of centralised remote control. Reports of swarm attacks on a Russian military base in Syria, for example, were based on the fact that the attackers used several small drones flying towards the target at the same time – but the drones weren’t collaborating or co-ordinating with each other.Wright, Tim, ‘When is a Drone Swarm Not a Swarm?’, Air and Space Magazine, 12 January 2018, https://www.airspacemag.com/daily-planet/when-drone-swarm-not-swarm-180967820/ On the other hand, some demonstrations of swarm technology are credible, indicating that true swarms may become a practical reality within the next 5-10 years.

Perhaps the most convincing demonstration to date of a true swarm comes from the US Air Force's Perdix drone project. A test in 2017 saw 103 of these small, low-cost drones being launched from F/A-18 Super Hornet aircraft and autonomously carrying out a series of manoeuvres.Mehta, Aaron, ‘Pentagon Launches 103 Unit Drone Swarm’, Defense News, 10 January 2017, https://www.defensenews.com/air/2017/01/10/pentagon-launches-103-unit-drone-swarm/ The drones themselves were designed using widely commercially-available components at MIT’s Lincoln Lab, carrying a small camera to capture images of targets. The Indian military’s more recent demonstration of a 75-strong swarm also seems broadly credible: this involved a group made up of several types of drone collaborating to identify and strike targets, although the degree to which the behaviour was pre-programmed is unclear, and the task was deliberately simple, using an uncluttered test area and stationary targets.Hambling, David, ‘Indian Army shows off drone swarm of mass destruction’, Forbes, 19 January 2021, https://www.forbes.com/sites/davidhambling/2021/01/19/indian-army-shows-off-drone-swarm-of-mass-destruction/?sh=2c9ec92e2384

The US’s success with Perdix reflects its position as the world leader in carrying out swarming research. China also produces a large volume of research in the area – ranking second globally in the count of scientific papers in the top 10% by citations (a measure that attempts to combine both volume and quality). But the quality of China’s output is difficult to judge: the standard metrics for research impact place it behind the UK and the US, and it is significantly less internationally-connected, with just 25% of Chinese swarming research involving international collaborations, compared with 60% for the UK or 40% for the US.

Military demonstrations and academic research aren’t the full story of swarming technology, however, even if they are the most eye-catching: there’s a lot of privately-funded development underway, too. Swarming systems have potential applications in several markets, drawing on their ability to multiply the capability a single human operator can wield: to plant seeds, spray fertilisers, and monitor crops in agricultural settingsWilson, Ken, ‘Swarms are coming to a farm near you’, Farm Weekly, 14 August 2019, https://www.farmweekly.com.au/story/6327157/swarms-are-coming-to-a-farm-near-you/ ; to map mining operationsStringer, David, ‘How Robots, Drones Are Transforming Mining and Mine Safety’, Insurance Journal, 4 April 2014, https://www.insurancejournal.com/news/national/2014/04/04/325475.htm ; to tackle wildfiresLenz, Max, ‘Drone Swarms for Firefighting: the Future of Fire Suppression?’, Drone Life, 28 April 2021, https://dronelife.com/2021/04/28/drone-swarms-for-firefighting-the-future-of-fire-supression/ ; or to survey marine environmentsDuarte, Miguel et al., ‘Application of swarm robotics systems to marine environmental monitoring,’ OCEANS 2016 - Shanghai, 2016, 10-13 April 2016, https://ieeexplore.ieee.org/document/7485429. Many of these applications involve relatively well-controlled environments, with limited amounts of complicating factors: the more the environment can be controlled, or at least well described, the easier it will be to use a swarm.