07Aerial robots, swarms & dual use

How to keep a robot fleet operating through communication and positioning failures

What keeps working, what degrades and when the mission must stop if GNSS, links or a fleet node fail. Every figure links to its source; where there is no sufficient public evidence, we say so.

Last reviewed: 2026-11Methodology & sources →

01 · The operational problem

“I need to inspect a facility even where coverage or positioning is unreliable”

Tunnels, mines, boiler interiors, substations and solar plants spanning hundreds of hectares share one problem: GNSS and radio are not guaranteed. GNSS interference is already documented by aviation authorities in several European regions.[EASA, 2026]

Operating envelope for each failure

No technique makes a fleet “work when everything fails”. What you can design is what the system does on each failure. This logic should be written into the operating procedure and the authorisation.

Failure	The mission continues if…	It degrades…	It must stop when…
GNSS loss or spoofing[EASA, 2026][IEEE ICRA, 2018]	A validated alternative localization source exists (VIO, LiDAR, beacons) and the environment is within its envelope (light, texture, geometry).	Position uncertainty grows with distance: reduce speed and range and prioritise the return.	There is no alternative source, or uncertainty exceeds the margin to obstacles or the authorised geofence.
Loss of the command link (C2)[EASA][IETF, 2022]	The authorised lost-link procedure allows completing the pre-programmed task; data is stored onboard.	No live video or re-tasking: the operator loses the ability to intervene.	As set by the contingency procedure: return to launch, to the dock, or a safe landing.
Network partition or leader loss[USENIX ATC, 2014]	The group that keeps a majority of nodes can elect a new leader and reassign tasks.	The minority group cannot elect a leader: it keeps already-assigned tasks that are safe without coordination.	A task requires coordination (several robots in the same volume, for example): the minority group suspends it.
Perception degradation (darkness, fog, dust)[IEEE ICRA, 2018]	Another sensor covers the condition: LiDAR in darkness, thermal in low light, or onboard lighting.	Obstacle avoidance and visual localization lose reliability: increase safety distances.	No onboard sensor ensures obstacle detection in the current condition.

GNSS loss or spoofing
The mission continues if…
A validated alternative localization source exists (VIO, LiDAR, beacons) and the environment is within its envelope (light, texture, geometry).
It degrades…
Position uncertainty grows with distance: reduce speed and range and prioritise the return.
It must stop when…
There is no alternative source, or uncertainty exceeds the margin to obstacles or the authorised geofence.
[EASA, 2026][IEEE ICRA, 2018]
Loss of the command link (C2)
The mission continues if…
The authorised lost-link procedure allows completing the pre-programmed task; data is stored onboard.
It degrades…
No live video or re-tasking: the operator loses the ability to intervene.
It must stop when…
As set by the contingency procedure: return to launch, to the dock, or a safe landing.
[EASA][IETF, 2022]
Network partition or leader loss
The mission continues if…
The group that keeps a majority of nodes can elect a new leader and reassign tasks.
It degrades…
The minority group cannot elect a leader: it keeps already-assigned tasks that are safe without coordination.
It must stop when…
A task requires coordination (several robots in the same volume, for example): the minority group suspends it.
[USENIX ATC, 2014]
Perception degradation (darkness, fog, dust)
The mission continues if…
Another sensor covers the condition: LiDAR in darkness, thermal in low light, or onboard lighting.
It degrades…
Obstacle avoidance and visual localization lose reliability: increase safety distances.
It must stop when…
No onboard sensor ensures obstacle detection in the current condition.
[IEEE ICRA, 2018]

02 · Decision table

What to evaluate, what to ask the vendor and what to measure in the pilot

Start from the need, not the technology. For each typical need, this table summarises the evaluation criteria, the questions to ask the vendor and the metrics that belong in the pilot's acceptance criteria.

Need	What to evaluate	Questions for the vendor	What to measure in the pilot
Indoor and confined-space inspection	GNSS-free localization (LiDAR, VIO) Onboard lighting and darkness sensors Collision and obstacle tolerance Recovery after link loss	Localization results in an environment equivalent to ours? What does the system do if it loses localization mid-gallery? What is the real endurance with the payload we need?	Position error against surveyed control points Share of surface with valid data Pilot interventions per flight Inspection time vs. current method
Fleets over large sites	Actual coordination level (central, distributed or swarm) Communications coverage across the site Battery and dock management Supervision: robots per operator	Who decides task allocation and what happens if that element fails? What data do robots share with each other and how often? What regulatory authorisation do similar operations hold?	Hectares inspected per hour and per operator Valid vs. planned coverage Time to reassign a failed robot's task Confirmed findings vs. false positives
Continuity under disconnection	Which tasks are safe without a link and which are suspended Onboard storage and forwarding when contact returns Authorised contingency procedures Behavior under network partition	Can we see a demonstration with an induced link loss? What happens if the fleet splits into two groups without contact? How is what happened during disconnection logged?	Missions completed with induced disconnection Data recovered vs. captured Reconnection and sync time Incidents and geofence deviations
Economic viability	Actual uptime (system availability) Data quality for the decision being made Maintenance, batteries and replacement Cost per mission and per useful finding	What is the measured availability at customers with similar use? What does the subscription include and what is extra? What is the service life of batteries and sensors?	Cost per mission vs. current method System availability (%) Maintenance hours per operating hour Findings that lead to a maintenance action

Indoor and confined-space inspection
What to evaluate
- GNSS-free localization (LiDAR, VIO)
- Onboard lighting and darkness sensors
- Collision and obstacle tolerance
- Recovery after link loss
Questions for the vendor
- Localization results in an environment equivalent to ours?
- What does the system do if it loses localization mid-gallery?
- What is the real endurance with the payload we need?
What to measure in the pilot
- Position error against surveyed control points
- Share of surface with valid data
- Pilot interventions per flight
- Inspection time vs. current method
Fleets over large sites
What to evaluate
- Actual coordination level (central, distributed or swarm)
- Communications coverage across the site
- Battery and dock management
- Supervision: robots per operator
Questions for the vendor
- Who decides task allocation and what happens if that element fails?
- What data do robots share with each other and how often?
- What regulatory authorisation do similar operations hold?
What to measure in the pilot
- Hectares inspected per hour and per operator
- Valid vs. planned coverage
- Time to reassign a failed robot's task
- Confirmed findings vs. false positives
Continuity under disconnection
What to evaluate
- Which tasks are safe without a link and which are suspended
- Onboard storage and forwarding when contact returns
- Authorised contingency procedures
- Behavior under network partition
Questions for the vendor
- Can we see a demonstration with an induced link loss?
- What happens if the fleet splits into two groups without contact?
- How is what happened during disconnection logged?
What to measure in the pilot
- Missions completed with induced disconnection
- Data recovered vs. captured
- Reconnection and sync time
- Incidents and geofence deviations
Economic viability
What to evaluate
- Actual uptime (system availability)
- Data quality for the decision being made
- Maintenance, batteries and replacement
- Cost per mission and per useful finding
Questions for the vendor
- What is the measured availability at customers with similar use?
- What does the subscription include and what is extra?
- What is the service life of batteries and sensors?
What to measure in the pilot
- Cost per mission vs. current method
- System availability (%)
- Maintenance hours per operating hour
- Findings that lead to a maintenance action

03 · Coordination levels

Not every group of drones is a swarm

The difference is not how many robots fly, but who decides and what information they share. Each level behaves differently under the same failure.

1
Independent drones
Who decides
Each operator or drone, separately
What they share
Nothing between drones; only telemetry with its operator
If a node or link fails
Its area stays uncovered until a person reassigns it.
Maturity and example
Common practice in inspection.
2
Fleet coordinated by a controller
Who decides
A central planner (station, cloud or fleet software)
What they share
State, position, battery and progress with the controller
If a node or link fails
If the controller or link fails, each drone finishes its sector or returns; reassignment stops.
Maturity and example
Authorised commercial operation: multi-drone solar inspection with software-assigned sectors, and remote BVLOS docks.[vHive, 2025][Skydio]
3
Distributed coordination
Who decides
The robots themselves, via rules or consensus (leader election, auctions)
What they share
State, tasks and coverage map, peer-to-peer over a mesh
If a node or link fails
The others reassign the failed node's task if they keep quorum; an isolated minority group cannot.
Maturity and example
Mature techniques in distributed systems and ad-hoc networks; we found no documented civil inspection operation with this architecture.[USENIX ATC, 2014][IETF, 2014]
No documented civil operation
4
Cooperating swarm
Who decides
Collective behavior emerging from each robot's local rules
What they share
Detections and partial maps that change neighbors' actions (confirm a finding, fill a gap)
If a node or link fails
Designed to degrade gradually as robots are lost, though its behavior is hard to verify and certify.
Maturity and example
Mainly research; the reference literature points out limits to real-world application.[Swarm Intelligence, 2013]
No documented civil operation

A mesh network only carries data between robots. Collective intelligence exists when what one robot shares (a detection, a gap in the coverage map, its battery) changes what the others do.

04 · Where Physical AI acts

Perceive, reason, decide and act — on each robot and across the fleet

What turns a drone fleet into a Physical AI system is that it closes the loop between what it perceives and what it does, without waiting for a human at every step. That loop exists at two scales, and each degrades differently.

Stage	On each robot	Across the fleet
1Perceive	Detect obstacles and anomalies (thermal, visible, LiDAR) with models running onboard.[Skydio, 2023]	Fuse detections from several robots into a shared map.
2Reason	Estimate its own position and uncertainty, and judge whether a detection is reliable.	Prioritise which findings to confirm and which coverage gaps to fill.
3Decide	Replan the route, slow down or return.	Reassign tasks by position, battery and sensors.
4Act	Execute flight or locomotion control.	Dispatch new tasks and create work orders.

Example on a solar plant

1A drone detects a panel with an abnormal temperature in its thermal image.
2It estimates the finding's position and confidence and shares it with the fleet.
3The system sends the nearest drone with a higher-resolution camera to confirm it, without interrupting everyone else's coverage.
4If confirmed, a work order is created with the location and images.

If the network goes down, each drone keeps its local loop (navigate, avoid obstacles, log findings) and the fleet loop is suspended until contact returns.

05 · Simulator

Try the failures in a teaching model

The simulator shows the coordination logic of an inspection fleet under interference, leader loss and network partition. It is a simplified teaching model, not a reproduction of real protocols or radio propagation. Its assumptions are listed next to it.

Didactic model · simplified simulation14 UAVs · solar plant · fixed seed

Simulated time T+0 s

Scenario

No failures. Each drone sweeps its sector and transmits. Three thermal anomalies are hidden in the panels.

Coordination mode

Cooperative swarm. Leaderless mesh: neighbors share tasks through local auctions; any drone with a ground link acts as gateway.

Starts when you reach this section or press Start.

Drone with a path to deliver data
Drone without a path (stores data)
Safe mode (no quorum)
Lost drone
Elected leader
Ground station (GCS)
Cell with delivered data
Cell swept, data on board
Orphaned sector
Anomaly detected (?) / confirmed (!)
Diversion to confirm
Interference zone · Links blocked (partition)

Event log

Perception → reasoning → decision → action, plus network events.

No events yet.

What happened and why

The explanation appears when the run ends. You can speed up with ×4.

Same failure, four ways to coordinate · Normal operation

Each row is a complete, deterministic run of the selected scenario, up to 180 simulated seconds, with the same rules and seed. The highlighted row is the one on screen.

Same failure, four ways to coordinate · Normal operation
Mode	Final valid coverage	Mission time	Orphaned sectors at the end	Reallocation after failure	Confirmed anomalies	Detection → confirmation	Data held on board (peak)
Independent routes	100%	42.6 s	—	—	0/3 · 3 to manual review	—	0%
Central controller	100%	53 s	—	—	3/3	5 s	0%
Distributed coordination	100%	53 s	—	—	3/3	5 s	0%
Cooperative swarm● On screen	100%	53 s	—	—	3/3	5 s	0%

Independent routes
Final valid coverage
100%
Mission time
42.6 s
Orphaned sectors at the end
—
Reallocation after failure
—
Confirmed anomalies
0/3 · 3 to manual review
Detection → confirmation
—
Data held on board (peak)
0%
Central controller
Final valid coverage
100%
Mission time
53 s
Orphaned sectors at the end
—
Reallocation after failure
—
Confirmed anomalies
3/3
Detection → confirmation
5 s
Data held on board (peak)
0%
Distributed coordination
Final valid coverage
100%
Mission time
53 s
Orphaned sectors at the end
—
Reallocation after failure
—
Confirmed anomalies
3/3
Detection → confirmation
5 s
Data held on board (peak)
0%
Cooperative swarm● On screen
Final valid coverage
100%
Mission time
53 s
Orphaned sectors at the end
—
Reallocation after failure
—
Confirmed anomalies
3/3
Detection → confirmation
5 s
Data held on board (peak)
0%

Times in simulated seconds, not real time. “—”: not applicable in this scenario. If you move the interference zone, the on-screen run diverges from this table.

Model assumptions

What this simulation represents and what it leaves out. Read this before drawing conclusions from the numbers.

Nature: Deterministic didactic model (fixed seed, 0.05 s step). It does not reproduce real protocols, radio propagation or flight dynamics, and it is not from an experiment. Its purpose is to compare coordination logics under the same rules.
Connectivity: Share of operational drones with a path to deliver data. Independent and central: direct link to the station. Distributed: one- or multi-hop mesh path to the leader. Swarm: mesh path to any drone with a ground link.
Coverage: We separate area swept (cell overflown) from valid coverage (its data reached the station). A drone without a path stores data on board and forwards it when it reconnects; once its task is done, it heads toward the station to deliver.
Links & interference: Link quality falls with the square of distance (mesh range 300 u., station 900 u.). Interference is a simplified factor: full strength in the inner 60 % of the zone, fading to the edge. Frequency hopping reduces it to 45 %. There is no radio physics.
Leadership: Raft-like election: each drone has a random timeout (1.5–3 s). When it expires without hearing from the leader, the drone runs as candidate and wins if its connected component holds a majority: 8 of the 14 configured members, failed ones included. A leader that loses the majority for 2 s steps down.
No quorum: A minority cannot elect a leader. It enters safe mode: it finishes its already-assigned task, stores data, accepts no new tasks and waits. This prevents two leaders, at the cost of not reacting to the unexpected.
Reallocation: A lost drone is detected after 3 s without heartbeats. The station reallocates in central mode, the leader in distributed mode, and neighbors after a 0.6 s auction in swarm mode. The lowest-cost drone wins: distance × (1 + 0.5 × pending tasks). The sector is redone in full.
Physical AI loop: Perception: when flying over a hot spot, a drone detects it with a fixed confidence (64–78 %). Reasoning: below 90 %, a second pass by another drone is required. Decision: the coordinator picks the lowest-cost drone. Action: that drone diverts, confirms (≥ 96 %) and returns to its lane. Without coordination, the anomaly is left for manual review.
Out of scope: Wind, battery physics and endurance, collisions and separation, real radio propagation, network latency, sensor errors, regulation (BVLOS, U-space) and security. The figures are not performance predictions for any real system.

06 · GNSS-denied navigation

Without satellites, the robot has to estimate its own position

Every localization source has an envelope: conditions where it works and conditions where it fails. Published figures come from benchmarks with specific hardware and environments; they are not performance guarantees at your facility.

Localization sources compared

Source	Published accuracy	Drift behavior	Where it fails	Compute
GNSS / RTK	~4.9 m (smartphone, open sky) · 1 cm + 1 ppm horizontal with fixed RTK (DJI M350 RTK sheet)[GPS.gov][Norsat][EASA, 2026]	No drift accumulation while the signal is valid	Jamming and spoofing, indoors, tunnels, dense urban areas and under foliage	Low
Inertial only (IMU)	No absolute accuracy: error grows with time squared (accelerometer bias) or cubed (gyro bias)[Inside GNSS]	Fast without external aiding	Any long mission without another source to bound the error	Low
Visual-inertial odometry (VIO)	Expressed as % of distance traveled; varies widely by algorithm, onboard computer and environment[IEEE ICRA, 2018][ETH Zürich, 2016]No verifiable general range	Grows with distance; loop closure reduces it at the cost of more compute	Low texture, darkness, fog, very fast motion and reflective surfaces	Medium
LiDAR SLAM	~0.5% translational error for top KITTI entries (ground vehicle outdoors)[KITTI]	Low; reduced further with loop closure or a prior map	Poor geometry (long uniform galleries), dense dust or rain; adds weight and power draw	High
Visual localization against satellite maps	Absolute; depends on altitude, terrain texture, season and map age[arXiv, 2025][arXiv, 2021]No verifiable general range	No drift accumulation when matching succeeds	Outdated maps, clouds, night without thermal, uniform terrain (water, desert) and indoors	High
Terrain-referenced navigation (TERCOM-style)	Limited by relief-map quality and by the terrain itself[USPTO]No verifiable general range	Absolute over terrain with relief	Flat terrain; needs an altimeter and an onboard digital elevation model	Low
UWB beacons	~10 cm with line of sight and calibration[NiceRF]	No drift accumulation within coverage	Needs installed infrastructure; multipath and poor anchor geometry	Low

GNSS / RTK
Low
Published accuracy
~4.9 m (smartphone, open sky) · 1 cm + 1 ppm horizontal with fixed RTK (DJI M350 RTK sheet)[GPS.gov][Norsat][EASA, 2026]
Drift behavior
No drift accumulation while the signal is valid
Where it fails
Jamming and spoofing, indoors, tunnels, dense urban areas and under foliage
Inertial only (IMU)
Low
Published accuracy
No absolute accuracy: error grows with time squared (accelerometer bias) or cubed (gyro bias)[Inside GNSS]
Drift behavior
Fast without external aiding
Where it fails
Any long mission without another source to bound the error
Visual-inertial odometry (VIO)
Medium
Published accuracy
Expressed as % of distance traveled; varies widely by algorithm, onboard computer and environment[IEEE ICRA, 2018][ETH Zürich, 2016]No verifiable general range
Drift behavior
Grows with distance; loop closure reduces it at the cost of more compute
Where it fails
Low texture, darkness, fog, very fast motion and reflective surfaces
LiDAR SLAM
High
Published accuracy
~0.5% translational error for top KITTI entries (ground vehicle outdoors)[KITTI]
Drift behavior
Low; reduced further with loop closure or a prior map
Where it fails
Poor geometry (long uniform galleries), dense dust or rain; adds weight and power draw
Visual localization against satellite maps
High
Published accuracy
Absolute; depends on altitude, terrain texture, season and map age[arXiv, 2025][arXiv, 2021]No verifiable general range
Drift behavior
No drift accumulation when matching succeeds
Where it fails
Outdated maps, clouds, night without thermal, uniform terrain (water, desert) and indoors
Terrain-referenced navigation (TERCOM-style)
Low
Published accuracy
Limited by relief-map quality and by the terrain itself[USPTO]No verifiable general range
Drift behavior
Absolute over terrain with relief
Where it fails
Flat terrain; needs an altimeter and an onboard digital elevation model
UWB beacons
Low
Published accuracy
~10 cm with line of sight and calibration[NiceRF]
Drift behavior
No drift accumulation within coverage
Where it fails
Needs installed infrastructure; multipath and poor anchor geometry

Figures depend on hardware, calibration, altitude, light and environment. Ask vendors for results under conditions equivalent to yours.

Teaching model

Position-estimate drift in a GNSS-denied gallery

A loop of about 520 m. Move the time slider to see how each estimate separates from the true path.

True path
IMU only (dead reckoning)49.3 m
VIO7.3 m
VIO + loop closure0.7 m

Mission time300 s

Model assumptions

Commercial-grade IMU with constant bias: error grows non-linearly with time (integrator physics, not data from a specific unit).
VIO with an assumed drift of 1.5% of distance traveled. An illustrative value, not a measured result.
Loop closure when revisiting known areas, pulling the estimate back near the true path.
No occlusions, no dust and enough texture: in real conditions, VIO can fail entirely.

Teaching model with editorial assumptions (evidence level D). It shows the shape of drift; do not use it to size a system.

07 · Resilient communications

Three layers that reduce the impact of a failure, each with its limits

Communication resilience combines the spectrum, the network topology and the decision logic. No layer removes the risk on its own: know what each one protects against and where it stops helping.

L1 · ESPECTRO
Spectrum: frequency hopping and selection
The link changes channel following a shared sequence and avoids bands where it detects noise.
Protects against
- Narrowband interference
- Momentary congestion of a channel
Limits
- Does not protect against interference covering all usable bands
- Available bands depend on national regulation
L2 · MANET
Topology: multi-hop ad-hoc network
Every robot also acts as a relay. Standardised routing protocols such as OLSRv2 recompute routes when a link drops; data can be stored onboard and forwarded when contact returns.[IETF, 2014][IETF, 2022]
Protects against
- Loss of individual links
- Robots beyond the station's direct range
Limits
- Reduces single points of failure but does not remove them: if a bridge node fails, the network splits
- Recovery time depends on the control-message configuration
- Store-and-forward does not guarantee delivery
L3 · CONSENSO
Decision: leader and tasks by consensus
With consensus algorithms such as Raft, a candidate needs a majority of votes to become leader. If the network splits, the minority group cannot elect a leader: a deliberate safety limit that prevents two leaders issuing contradictory orders.[USENIX ATC, 2014]
Protects against
- Leader failure while a majority remains
- Contradictory orders after a partition
Limits
- A partition without a majority halts coordination in that group
- Tasks must be designed to be safe without coordination

08 · Operational lessons & civil maturity

What hostile environments have taught, and what is already proven in civil use

Recent conflicts have accelerated technologies the civil sector can use. We separate two things: the publicly documented lesson and the demonstrated maturity of a specific civil system. Where no civil system has documented operation, we say so.

Key transfer

Multi-drone coordinationSolar farms & forestsTRL 9
GNSS-free navigation (VIO/SLAM)Underground mining & tunnelsTRL 9
Low-latency first-person (FPV) pilotingDisaster search & rescueNo sufficient public evidence

Context

Civil application

7/7 technologies

Multi-drone coordination
★ Key transfer: Solar farms & forests
Several drones sharing a mission, with task allocation and reassignment on failure.
USASolar farms & forestsAgricultural monitoringDisaster search & rescue
Demonstrated civil maturity
Demonstrated civil maturity: TRL 9
Multi-drone solar inspection with software-assigned sectors (vHive, FAA waiver, 2025). This is central coordination, not a decentralised swarm.[vHive, 2025]
Decentralised swarms in civil inspection: we found no documented operation.
GNSS-free navigation (VIO/SLAM)
★ Key transfer: Underground mining & tunnels
Estimating position from cameras, IMU and LiDAR when satellites are unavailable or unreliable.
RussiaUnderground mining & tunnelsCritical infrastructure
Demonstrated civil maturity
Demonstrated civil maturity: TRL 9
LiDAR SLAM drone in a coal mine (Elios 3 at LW Bogdanka), documented in a technical publication.[E3S Web of Conferences, 2024]
Low-latency first-person (FPV) piloting
★ Key transfer: Disaster search & rescue
Agile manual control with low-latency video to maneuver in complex spaces.
UkraineDisaster search & rescueCritical infrastructureUnderground mining & tunnels
Demonstrated civil maturity
Demonstrated civil maturity: TRL 9
Assisted piloting in confined spaces with a caged drone (Elios 3 in a mine).[E3S Web of Conferences, 2024]
FPV use in search and rescue: we found no public evaluation of results.
Fiber-optic control links
The drone receives commands over an unspooling fiber cable: it emits no radio and is unaffected by jamming.
UkraineRussiaUnderground mining & tunnelsCritical infrastructure
Demonstrated civil maturity
No sufficient public evidence
We found no documented civil deployments with unspooling fiber. Tethered drones are a different technology.
Visual localization against maps
Matching the camera image against onboard satellite maps to obtain an absolute position without GNSS.
RussiaAgricultural monitoringDisaster search & rescueLong-range logistics
Demonstrated civil maturity
No sufficient public evidence
Recent literature is mostly research prototypes; we found no documented civil operation.
Mass production and fast iteration cycles
Simple designs built in volume with commercial components and frequently updated software.
IranUSALong-range logisticsAgricultural monitoringCritical infrastructure
Demonstrated civil maturity
Demonstrated civil maturity: TRL 9
Civil autonomous operation at scale: medical and retail logistics with over 2 million deliveries (Zipline, company figures). It does not derive from these programs.[DroneLife, 2026]
Docked drones and remote BVLOS operation
Autonomous stations that launch, recharge and recover the drone, operated remotely.
USAPerimeter securityCritical infrastructureDisaster search & rescue
Demonstrated civil maturity
Demonstrated civil maturity: TRL 9
BVLOS docks at dozens of a utility's facilities (Skydio with Dominion Energy, 2022, per the manufacturer).[Skydio]

TRL is assigned to a specific system, not to a technology category, using NASA's scale (9 = actual system proven through successful operations). The assignment is ours, based on the cited evidence.[NASA]

High-level analysis from public sources, focused on civil applications. It contains no operational or employment detail.

09 · Regulatory framework (EU)

Which regulatory route applies to your operation

In the EU, an operation's category depends on its risk. Most professional BVLOS inspections fall into the specific category, which has several access routes with different requirements.

UE 2019/947
Three risk-based categories
Open: low risk, with height, weight and visual-line-of-sight (VLOS) limits; no prior authorisation. Specific: higher-risk operations, such as most professional BVLOS. Certified: risk comparable to crewed aviation (e.g., carrying people).[EUR-Lex, 2019]
STS
Standard scenario: declaration
If the operation fits an EASA standard scenario (STS-01 and STS-02), a declaration to the authority is enough, with no prior authorisation. The scenario's conditions and the matching equipment class are required.[EASA][Transpordiamet]
PDRA · SORA
Authorisation: PDRA or SORA
If it does not fit an STS, an authorisation is requested. A PDRA is a scenario already assessed by EASA with defined mitigations, simplifying the application. If no PDRA fits either, a full SORA risk assessment is done; its risk level drives the link, containment and contingency-procedure requirements.[EASA]
LUC
LUC certificate: self-approval
The light UAS operator certificate (LUC) lets an organisation with a mature management system approve its own operations within the granted privileges. It is the route to scale many operations without requesting an authorisation for each one.[EASA][Luftfartstilsynet]
UE 2021/664
U-space
Framework to designate U-space airspace with digital services for drones, such as network identification, geo-awareness and flight authorisation. Its rollout depends on each state designating such airspace.[EUR-Lex, 2021]
UE 2024/1689
AI Act and aviation
For civil aviation, the AI Act acts mainly through EASA rules: aviation acts on AI systems that are safety components must take the high-risk requirements into account. Exclusively military systems are outside its scope.[EASA, 2024][SpringLex]
UE 2021/821
Dual-use exports
Items on the Annex I list need a licence for export outside the EU. Check the classification of navigation, communications and encryption hardware and software before selling or transferring.[EUR-Lex, 2021]

Informational summary as of November 2026; not legal advice. Always check the consolidated text and your national authority.

10 · Platforms

Aerial platforms related to these applications

Open a sheet to see specifications with their conditions and sources.

Explore all platforms →

Dual-use ground platforms (e.g., THeMIS or Vision 60) are in the explorer, under ground vehicles and quadrupeds. See ground platforms →

Sources for this page

How we grade evidence →

1.
EASA updates Safety Information Bulletin on GNSS interference (SIB 2022-02R4) — EASA, 2026-07
A· Documentedaccessed 2026-11-02
GNSS jamming and spoofing have increased since February 2022; affected areas include the Black Sea, Baltic, Mediterranean, Middle East and Arctic.
2.
A Benchmark Comparison of Monocular Visual-Inertial Odometry Algorithms for Flying Robots (Delmerico & Scaramuzza, Univ. of Zurich) — IEEE ICRA, 2018-05
A· Documentedaccessed 2026-11-02
Compares six VIO algorithms on EuRoC and onboard computers: accuracy and compute cost vary widely by algorithm and hardware; loop closure improves accuracy at the cost of more resources.
3.
EU UAS Regulation — specific category — EASA
A· Documentedaccessed 2026-11-02
Specific-category routes: declaration under a standard scenario (STS), authorisation via PDRA or SORA, and the LUC certificate.
4.
RFC 9171 — Bundle Protocol Version 7 — IETF, 2022-01
A· Documentedaccessed 2026-11-02
Store-carry-forward overlay for intermittent links; the protocol alone does not guarantee delivery.
5.
In Search of an Understandable Consensus Algorithm — Raft (Ongaro & Ousterhout, Stanford) — USENIX ATC, 2014
A· Documentedaccessed 2026-11-02
A candidate needs a majority of votes to become leader; the system is only available while a majority of nodes can communicate.
6.
vHive Exclusively Authorized by FAA for its Autonomous Multi-Drone Solar Inspections in the US — vHive, 2025-01
B· Manufacturer / operatoraccessed 2026-11-02
FAA waiver for solar inspection with several drones at once; the software splits the site into sectors and assigns one per drone.
7.
Regulatory scaling of remote drone operations — Skydio
B· Manufacturer / operatoraccessed 2026-11-02
Per Skydio: FAA waiver (2022) with Dominion Energy for BVLOS operations at dozens of facilities in six states without an additional visual observer.
8.
RFC 7181 — The Optimized Link State Routing Protocol Version 2 (OLSRv2) — IETF, 2014-04
A· Documentedaccessed 2026-11-02
9.
Swarm robotics: a review from the swarm engineering perspective (Brambilla, Ferrante, Birattari & Dorigo; vol. 7, 1–41) — Swarm Intelligence, 2013-01
A· Documentedaccessed 2026-11-02
Reference review; includes a section on the current limits of swarm robotics as an applied engineering discipline.
10.
Skydio X10 and the Next Chapter of Drones in Critical Industries — Skydio, 2023
B· Manufacturer / operatoraccessed 2026-11-02
Visual navigation with six cameras, onboard NVIDIA Jetson Orin compute and night flight (NightSense). Capabilities as stated by the manufacturer.
11.
GPS Accuracy — GPS.gov
A· Documentedaccessed 2026-11-02
GPS-enabled smartphones: typically ~4.9 m radius under open sky; degrades near buildings, bridges and trees.
12.
Matrice 350 RTK (NA) SP Combo — specifications — Norsat
C· Press / analysisaccessed 2026-11-02
Reproduces the manufacturer sheet: RTK FIX 1 cm + 1 ppm (horizontal), 1.5 cm + 1 ppm (vertical). Only with a fixed RTK solution and corrections available.
13.
Inertial Error Propagation: Understanding Inertial Behavior — Inside GNSS
C· Press / analysisaccessed 2026-11-02
A constant accelerometer bias yields position error growing with time squared; gyro bias, with time cubed in the short term.
14.
The EuRoC micro aerial vehicle datasets (Burri et al., IJRR) — ETH Zürich, 2016
A· Documentedaccessed 2026-11-02
Reference dataset for drone VIO, with laser-tracker and Vicon ground truth. Indoor scenarios.
15.
Visual Odometry / SLAM Evaluation 2012 — KITTI (KIT & TTI-C)
A· Documentedaccessed 2026-11-02
Mean translational error over 100–800 m subsequences, in %. Top entries are around 0.5%. A ground-vehicle outdoor benchmark: not directly transferable to drones.
16.
Hierarchical Image Matching for UAV Absolute Visual Localization via Semantic and Structural Constraints — arXiv (preprint), 2025
C· Press / analysisaccessed 2026-11-02
Preprint. Absolute localization by matching a downward camera against satellite imagery; viewpoint, scale and seasonal differences are the main difficulty.
17.
GNSS-denied geolocalization of UAVs by visual matching of onboard camera images with orthophotos (Kinnari et al.) — arXiv, 2021
C· Press / analysisaccessed 2026-11-02
Combines VIO, orthorectification and Monte Carlo localization against orthophotos.
18.
Systems and methods for a terrain contour matching navigation system (US7522090B2) — USPTO
B· Manufacturer / operatoraccessed 2026-11-02
Describes the principle: matching altimeter profiles against a stored relief map to correct inertial drift. A patent does not validate performance.
19.
DW3000 UWB Indoor Positioning: TWR, Anchors & Calibration — NiceRF
B· Manufacturer / operatoraccessed 2026-11-02
~10 cm ranging accuracy with line of sight and calibration; worse with multipath and poor anchor geometry.
20.
Replicator — Defense Innovation Unit
A· Documentedaccessed 2026-11-02
Program's stated goal (August 2023): low-cost autonomous systems at the scale of thousands within 18–24 months.
21.
DOD Replicator Initiative: Background and Issues for Congress (IF12611) — Congressional Research Service
A· Documentedaccessed 2026-11-02
The goal was thousands of systems by August 2025; per a former official cited by CRS, hundreds had been fielded.
22.
Finnair suspends flights to Estonian city over Russian GPS interference — AFP, 2024-04
C· Press / analysisaccessed 2026-11-02
Finnair suspended the Helsinki–Tartu route after two flights lost GNSS signals.
23.
Baltic countries blame Russia for GPS jamming of commercial flights — The Record, 2024-04
C· Press / analysisaccessed 2026-11-02
24.
The practical use of UAV systems in mine excavations and inspection works at LW “Bogdanka” S.A. (Szadziul et al.; vol. 567, 01009) — E3S Web of Conferences, 2024-09
A· Documentedaccessed 2026-11-02
Elios 3 use in a coal mine: storage-tank inspection, dewatering roadways and measurements in hard-to-reach areas. Written by the mine's own staff.
25.
Production of Ukrainian drones. The Ministry of Defense reported on the over-implementation of the 2024 plan — Online.ua, 2025
C· Press / analysisaccessed 2026-11-02
Ukraine's Ministry of Defense said it exceeded the 2024 plan with over one million units.
26.
Ukraine and Russia deploy next-level fibre optic drones which avoid electronic jamming — Forces News
C· Press / analysisaccessed 2026-11-02
27.
Ukraine is buying jam-proof fiber-optic drones for first time to single standard, so any maker can compete — Euromaidan Press, 2026-08
C· Press / analysisaccessed 2026-11-02
28.
Iran’s Shahed drone: How the ‘poor man’s cruise missile’ is shaping Tehran’s retaliation — The Nightly
C· Press / analysisaccessed 2026-11-02
29.
DIU Overhauls Blue UAS Program to Accelerate Drone Integration Across the Department of Defense — Inside Unmanned Systems, 2025
C· Press / analysisaccessed 2026-11-02
30.
Zipline surpasses 2 million deliveries with expansion to Houston and Phoenix — DroneLife, 2026-01
C· Press / analysisaccessed 2026-11-02
Company-reported figures: over 2 million commercial deliveries and over 125 million autonomous miles.
31.
Technology Readiness Levels — NASA
A· Documentedaccessed 2026-11-02
1–9 scale; TRL 9 = actual system proven through successful operations.
32.
Commission Implementing Regulation (EU) 2019/947 on the rules and procedures for the operation of unmanned aircraft — EUR-Lex, 2019-05
A· Documentedaccessed 2026-11-02
33.
Specific category flights — Transpordiamet (Estonia)
A· Documentedaccessed 2026-11-02
EASA standard scenarios: STS-01 and STS-02. National specific-category authorisations are valid in other EASA states.
34.
Specific category — Luftfartstilsynet (Noruega)
A· Documentedaccessed 2026-11-02
A LUC holder may be entitled to approve its own SORA, PDRA or STS operations.
35.
Commission Implementing Regulation (EU) 2021/664 on a regulatory framework for the U-space — EUR-Lex, 2021-04
A· Documentedaccessed 2026-11-02
36.
Regulation (EU) 2024/1689 of the European Parliament and of the Council (AI Act) — EASA, 2024-07
A· Documentedaccessed 2026-11-02
37.
AI Act, Article 108 — Amendments to Regulation (EU) 2018/1139 — SpringLex
C· Press / analysisaccessed 2026-11-02
Reproduction of the official text. Aviation delegated and implementing acts on AI systems that are safety components must take the AI Act's high-risk requirements into account.
38.
Regulation (EU) 2021/821 — Union regime for the control of exports, brokering, technical assistance, transit and transfer of dual-use items (recast) — EUR-Lex, 2021-05
A· Documentedaccessed 2026-11-02

How to keep a robot fleet operating through communication and positioning failures

Operating envelope for each failure

GNSS loss or spoofing

Loss of the command link (C2)

Network partition or leader loss

Perception degradation (darkness, fog, dust)

What to evaluate, what to ask the vendor and what to measure in the pilot

Indoor and confined-space inspection

Fleets over large sites

Continuity under disconnection

Economic viability

Not every group of drones is a swarm

Independent drones

Fleet coordinated by a controller

Distributed coordination

Cooperating swarm

Perceive, reason, decide and act — on each robot and across the fleet

Example on a solar plant

Try the failures in a teaching model

Event log

What happened and why

Same failure, four ways to coordinate · Normal operation

Without satellites, the robot has to estimate its own position

Localization sources compared

GNSS / RTK

Inertial only (IMU)

Visual-inertial odometry (VIO)

LiDAR SLAM

Visual localization against satellite maps

Terrain-referenced navigation (TERCOM-style)

UWB beacons

Position-estimate drift in a GNSS-denied gallery

Three layers that reduce the impact of a failure, each with its limits

Spectrum: frequency hopping and selection

Topology: multi-hop ad-hoc network

Decision: leader and tasks by consensus

What hostile environments have taught, and what is already proven in civil use

Multi-drone coordination

GNSS-free navigation (VIO/SLAM)

Low-latency first-person (FPV) piloting

Fiber-optic control links

Visual localization against maps

Mass production and fast iteration cycles

Docked drones and remote BVLOS operation

Which regulatory route applies to your operation

Three risk-based categories

Standard scenario: declaration

Authorisation: PDRA or SORA

LUC certificate: self-approval

U-space

AI Act and aviation

Dual-use exports

Aerial platforms related to these applications

X10

Matrice 350 RTK

Elios 3

Zip (Platform 1)

Vector