Swarm robotics: how groups of simple robots work together

Swarm robotics is an approach in which many small, simple robots work as a group without a central controller. Each robot senses only its nearby neighbours and follows simple local rules, and the group's overall behaviour comes from those many small interactions. It matters because a group can keep working when some of its members fail, and because cheap units could, in principle, cover large areas. The public evidence is still mostly from laboratories and short trials. The large drone shows and warehouse fleets that people hear about are not clearly decentralised: their public descriptions do not show robots deciding for themselves.

What a swarm is, compared with one robot

A single robot is usually one machine with a plan, a set of sensors and a controller that understands the whole task. A swarm reverses that idea. Each unit is deliberately limited. In August 2014, Harvard researchers reported using 1,024 small Kilobot robots, each of which could communicate with a few neighbours and had no overhead view of the group. The team sent commands to all the robots at once using infrared light, and the robots arranged themselves into shapes such as a five-pointed sea star and the letter K (Harvard SEAS, 2014). Four robots marked an origin, and the others tracked their distance from it and followed the edge of the group (Harvard SEAS, 2014).

The trade-off is clear. Any one unit is weak, so the reliability has to come from the group as a whole.

Demonstrated under test

These results were shown in trials. Each entry states its setting and size. None of them is a commercial deployment.

• Shape forming, 2014 (test, not a deployment). 1,024 Kilobots formed two-dimensional shapes. A Harvard researcher said the robots' abilities were "more variable and less reliable" than those of conventional robots. Errors such as traffic jams and off-course robots became much more common in a large group, and nearby robots sensed some of them and cooperated to fix them (Harvard SEAS, 2014). This summary is based on the university's press release rather than the paper in Science.
• Aquatic robots (peer-reviewed, PLoS ONE, 2016). Up to ten surface robots on water ran neural controllers that had been evolved in simulation. The tasks were homing, dispersion, clustering and monitoring. The authors report performance similar to simulation, in what they describe as a real and uncontrolled environment. The abstract gives no numerical results (Duarte et al., arXiv version of the PLoS ONE paper).
• Indoor drone search, 2019 (test, not a deployment). Researchers from TU Delft, Radboud University and the University of Liverpool flew six 33-gram camera drones in an indoor office containing two dummies that stood in for disaster victims. The team reported that the drones explored about 80 per cent of the open rooms in six minutes (Security Magazine, 2019). This is a secondary report of the work, so treat the figure with care.

Reported, but not independently checked

These claims come from companies, military officials, a government research service or the press. We could not verify them independently, and each label shows who made the claim.

• Warehouse fleets (company announcement). In July 2025 Amazon said it had deployed its one millionth robot, in a network spanning more than 300 facilities, and that a new model called DeepFleet would improve the travel time of its robotic fleet by 10 per cent (Amazon, 2025). The 10 per cent is Amazon's own claim, and the page does not show the data behind it. Amazon compares DeepFleet to an intelligent traffic management system that coordinates robot movement across its fulfilment network. The page does not say whether that coordination is central or distributed.
• Record drone flight (company flight, record listed by Guinness World Records). Guangdong EHang Egret Media Technology flew 22,580 drones over Hefei, China, on 3 February 2026. Guinness World Records lists the flight as the most drones airborne at once from a single computer, outdoors (Guinness World Records). The record title implies that one computer ran the flight, but the page does not describe how it was controlled.
• Sydney show failure (company statement, press). In May 2026, 89 of about 1,000 drones at a Vivid festival show in Sydney fell, 83 of them into Cockle Bay. The operator, SkyMagic, blamed radio interference. The Australian Transport Safety Bureau opened an investigation, and the report gave no findings (ABC News, 2026). In the same report, the head of another drone show company said each show drone flies a programmed flight path and is monitored from a master computer by a pilot.
• Military programme (official remarks, press). In August 2023 the US Department of Defense set a goal for its Replicator initiative to field systems at a scale of "multiple thousands" within 18 to 24 months (DefenseScoop, 2025). Several current and former officials suggested that hundreds had been delivered, and T.S. Allen, a former director of Replicator-1 at the Defense Innovation Unit, said thousands more were on contract. Defence officials did not verify these remarks (DefenseScoop, 2025). A Congressional Research Service report, updated on 21 January 2026, says swarming "would likely require further advancements in artificial intelligence and/or networked communications to be deployed". It also says the Department has released only limited information about specific Replicator capabilities or systems (CRS, 2026).
• Multi-drone approvals (company announcement). According to DroneXL, Skydio announced on 26 March 2026 that the US Federal Aviation Administration had approved 12 public safety agencies to let one pilot operate up to four Skydio X10 drones at once. The specific conditions of each waiver were not published with the announcement (DroneXL, 2026). A pilot supervising several drones is not the same as drones coordinating among themselves.

Speculative applications (proposals, no evidence of use)

These are proposals that follow from the research direction. We have found no evidence that any of them is in use.

• Search and monitoring of disaster areas by many low-cost robots that keep working when individual units are lost. The indoor test above does not show a deployed search system.
• Environmental monitoring of lakes or coastal waters by small groups of surface robots, building on the aquatic trial.
• Collective construction of simple structures by robots that build together without a central plan, extending the 2014 shape-forming work.

Limits and open problems

• Unreliable parts. Cheap robots struggle to move in a straight line, their distance sensing varies, and they get caught in traffic jams. Errors become more common as groups grow, so the design has to correct for failure rather than assume it away.
• Limited communication. Each robot knows only its neighbours. The global goal has to be built into local rules, and that makes the design harder to check.
• Battery life. The sources reviewed did not report how long these swarms can run on a charge, so we cannot say whether battery life limits current trials.
• Testing outside labs. The aquatic trial used up to ten robots, and the indoor drone test used six. The sources reviewed include no outdoor field trial of a decentralised swarm in 2025 or 2026.
• Safety and control. Airspace rules, unpublished waiver conditions, the Sydney failure and limited military disclosure all mean that the public cannot yet judge how these systems fail.

What to watch next

• Outdoor field results from swarms of a few dozen robots or more that coordinate without a central controller, published with failure counts.
• The published conditions of FAA waivers that allow one pilot to supervise several drones.
• Findings from the Australian Transport Safety Bureau on the Sydney show failure.
• Verified Replicator delivery figures, or an account of how those systems are controlled.
• Independent checks of Amazon's 10 per cent travel-time claim.