Autonomous drone flight for inspection & delivery

Manual drone piloting does not scale: every flight needs a pilot on site and an observer. Autonomous drones, often stationed in a weatherproof dock and flown beyond visual line of sight (BVLOS) from a remote control centre, fly repeatable inspection routes or carry small payloads (medical samples, spare parts, parcels). On-board AI handles navigation, obstacle avoidance and landing, while humans supervise by exception. The binding constraint in the EU is less the AI than the airworthiness and operational authorisation.

Problem & value

• Today: inspections of power lines, wind turbines, bridges, solar parks, roofs and stockpiles use rope access, helicopters or manually piloted drones. Deliveries of urgent small items (lab samples, parts between sites) go by road courier. Both are slow, costly and sometimes dangerous.
• AI value: repeatable, scheduled data capture without crew on site; fewer people at height; faster delivery between hospital sites or across water and terrain; consistent data for change detection (see VIS-003) and predictive maintenance (PRED-006).
• Quantified benefits vary by operator and are mostly vendor-reported; measure your own baseline (cost per inspected asset, delivery time) before committing.

Variants

• Docked inspection ("drone-in-a-box"): scheduled flights from a dock at a substation, plant, quarry or port; remote pilot supervises several docks.
• Linear infrastructure BVLOS: power lines, pipelines, railways, roads.
• Confined-space / indoor inspection: collision-tolerant drones in tanks, boilers, mines, where no GNSS exists.
• Light-payload delivery: hospital-to-hospital, offshore/island supply, rural parcels, industrial site logistics.
• Warehouse inventory flights: see VIS-019. Disaster mapping: see VIS-058.
• Private use: consumer drones with follow-me and obstacle avoidance are covered by the open category; autonomy beyond that is not a consumer use case.

How it works

1. mission request (asset list / delivery order)
2. planning (route optimisation, airspace + geo-zone check, weather, battery/energy budget)
3. authorisation (operational authorisation, U-space / flight plan, risk assessment)
4. autonomous flight (GNSS/RTK + visual-inertial odometry, SLAM, obstacle avoidance, detect-and-avoid)
5. mission payload (inspection capture | package release / winch)
6. precision landing / dock, battery swap or charge
7. cloud/edge analysis (defect detection, 3D model, change detection) | delivery confirmation
8. remote pilot review, work orders, flight log archive

• Perception: stereo/depth cameras or LiDAR with learned obstacle detection and visual-inertial odometry; deep-learning detectors for landing zones, people and vehicles below.
• Planning & control: classical planners (A*, sampling-based, model-predictive control) remain the safety-relevant core; learned policies (RL, sim-to-real, see ROBO-010) are used for agile flight and landing but rarely as the only safety layer.
• Detect-and-avoid for BVLOS: cooperative (ADS-B, FLARM, remote ID) plus non-cooperative sensing (radar, acoustic, vision); a certified-style safety monitor can override the AI.
• Human-in-the-loop: remote pilot with take-over authority, automatic return-to-home, parachute or flight-termination system; defect findings are confirmed by an inspector.

Data preparation

The data is inspection imagery (RGB, thermal, lidar), flight logs and maps, plus asset registers from the ERP, CMMS or GIS to which findings must be linked. Images typically sit on the drone, a vendor cloud or SD cards and file shares; asset master data is often incomplete or inconsistent. Typical gaps are missing georeferencing, no consistent asset IDs, unlabelled defects, duplicate captures and unclear ownership of vendor-held data.

Preparation steps: agree on asset IDs and a naming scheme (PREP-010, PREP-011), export asset data from ERP/CMMS (PREP-005), define defect categories before labelling (PREP-019), harvest labels from past inspection reports (PREP-020) and build a test set of annotated images with your inspectors (PREP-018). Photogrammetry tools such as OpenDroneMap or Pix4D prepare maps. Check export rights and formats in the vendor contract. Preparing the asset data and historical labels usually accounts for roughly a third of the effort; the flight system itself is largely vendor-supplied.

Implementation paths

Which path for whom

• SMB: Path A, because certified docked systems from EU vendors exist and most firms should buy flight autonomy and operator authorisation, not build them.
• Enterprise (utilities, large asset owners): Path B, because inspection results must flow into asset-management systems, so vendor APIs plus own analytics pay off at fleet scale.
• Public (municipalities, authorities): Path A, because certified products with a clear operator authorisation fit procurement rules; sovereign EU hosting for the imagery is a tender criterion.
• Developer: Path C, because PX4/ArduPilot, ROS 2 and open detectors give full control, with airworthiness evidence remaining your burden.

Path A — Buy (recommended for SMBs)

• Docked inspection platforms: Skydio (X10 with Dock, US), DJI Dock (China; check security policies of your organisation and customer), Percepto (Israel), Flyability Elios 3 for confined spaces (Switzerland, EU-adjacent), Parrot ANAFI (France, EU), Quantum Systems (Germany, EU), Wingtra (Switzerland) for mapping.
• Delivery: Wingcopter (Germany, EU), Matternet (US/Switzerland), Manna (Ireland, EU), Zipline (US), Wing (US).
• Fleet and mission software: DroneDeploy, Pix4D (Switzerland), Hexagon, Dronehub / UTM providers for U-space (e.g. Unifly, Belgium, EU).
• Choose A when: you need a certified, supported system and an operator with an authorisation; most organisations should buy flight autonomy and build only the analytics.

Path B — Build on cloud APIs

• Use vendor SDKs and open interfaces (Skydio Cloud API, DJI Cloud API, MAVLink-based autopilots) to trigger missions and pull media.
• Analytics in EU regions: managed vision services (Azure, Google, AWS) or a VLM for open-ended defect description over inspection imagery; orchestration via workflow tools.
• Keep the flight-critical loop on the aircraft, never in a cloud round trip.

Path C — Self-host (open source)

• Autopilot and middleware: PX4 and ArduPilot (MAVLink), ROS 2, Gazebo/AirSim-style simulators, MAVSDK, Aerostack2; VINS-Fusion, ORB-SLAM3, OpenVINS for state estimation; Ego-Planner style obstacle-free planning.
• On-board AI: YOLO-family detectors or RT-DETR exported to TensorRT on NVIDIA Jetson Orin or Qualcomm/Hailo modules (HW1); see ROBO-011 for edge deployment.
• Analytics: open-weight segmentation and open-vocabulary detectors, OpenDroneMap for photogrammetry.
• Effort: high; airworthiness evidence remains your burden.

Path D — Custom training

Worthwhile for defect classes unique to your assets (insulator damage, blade erosion, weld corrosion) — a few thousand labelled images per class, ideally augmented with synthetic data. For flight policies, train in simulation with domain randomisation and validate extensively in the real world.

State of the art (2026)

• Docked autonomy is production at leading utilities, mines and security-sensitive sites; one remote operator supervises several docks.
• BVLOS in the EU is increasingly granted via the specific category with SORA risk assessment and predefined scenarios; U-space services (network remote ID, geo-awareness, flight authorisation) are rolling out country by country.
• Learned vision-based flight has reached champion-level performance in racing research, but such policies are not yet certifiable for routine operations.
• Foundation models are being used for inspection reporting (VLM captions of defects) and open-vocabulary detection; hallucination makes human sign-off mandatory.
• Limits: wind and rain, GNSS jamming/spoofing, battery endurance, payload mass, low-contrast obstacles (wires, glass), regulatory lead times, and public acceptance of noise and privacy.

Shared resources & combinations

• Perception stack (object-detection, segmentation, 3d-vision, edge-runtime) is shared with ROBO-002 AMRs, ROBO-015 inspection robots, ROBO-020 sidewalk delivery and ROBO-005 driving.
• Simulation and sim-to-real shared with ROBO-010.
• Downstream: VIS-003 (damage detection from imagery), PRED-006 (maintenance prediction), PRED-017 (weather forecasts for go/no-go), VIS-058 (disaster mapping).
• Counter-drone: VIS-044 protects sites from non-cooperative drones and must be coordinated with your own flights.

Effort, cost & KPIs

• Path A: weeks to months, mainly authorisation, site survey and integration; hardware plus per-dock or per-flight subscription.
• Path B: 3-6 months for analytics integration.
• Path C: 12+ months and a specialised team, plus certification effort.
• Cost drivers: authorisation and insurance, remote-pilot staffing, hardware lifecycle, data labelling.
• KPIs: mission completion rate, interventions per 100 flights, cost per inspected asset or per delivery, defect recall/precision, delivery time, incident rate, availability under weather limits.

People & change

Management sets the scope and owns the aviation and liability risk. The department lead (asset or grid operations) and the inspectors who work with the images are involved from the first workshop: they define what a relevant defect is and validate results. IT security and IT cover cloud links, docking stations and access rights; legal/compliance handles drone operator authorisation and insurance. Data owners of asset data join at the data-preparation stage.

Works council: depends. Missions are normally not aimed at staff, but logging pilots' and inspectors' work and capturing people on site can be suited to monitor behaviour (§87(1) no. 6 BetrVG); §90(1) no. 3 requires early information when work processes change. Public bodies involve the Personalrat (GUIDE-004).

Data protection: involve the DPO early, because overflight imagery can capture people and property; a DPIA is likely for regular flights over populated areas (GUIDE-005). AI literacy: support staff competence with short training for inspectors on strengths and limits of automatic detection (GUIDE-006). Start with a pilot at one site, communicate that drones replace dangerous climbs, not people, and keep inspectors as the decision makers on findings and a remote pilot able to intervene (GUIDE-008).

Risks, regulation & compliance

• Aviation law: Regulation (EU) 2019/947 and 2019/945 define open, specific and certified categories; autonomous BVLOS usually needs specific-category authorisation (SORA, operator authorisation) and, for heavier or populated-area operations, certification. Remote ID and U-space (Reg. 2021/664) apply. National authorities (e.g. LBA/DFS in Germany) implement.
• EU AI Act: classification depends. Drones as such fall under aviation regulation that the AI Act only touches via delegated acts for Reg. 2018/1139 (Annex I Section B); AI used as a safety component of products under other Annex I law can become high-risk. Inspection analytics are typically minimal risk, unless used for critical infrastructure safety management (Annex III 2). Art. 5 prohibitions: drones with biometric identification or categorisation of people, or used for law-enforcement real-time remote biometric identification, are heavily restricted or prohibited; do not add such functions to inspection or delivery drones.
• GDPR: cameras capture bystanders, vehicles and private property; DPIA, minimised data, blurring of persons and licence plates, signage/information, short retention, purpose limitation (EDPB guidance on video devices applies).
• Security: GNSS spoofing, command-link hijack, supply-chain concerns for foreign hardware and cloud, NIS2 for operators in essential sectors, Cyber Resilience Act for products.
• Safety: ground-risk and air-risk mitigation, flight termination, insurance, noise and wildlife rules.
• Mitigations: conservative geofencing, redundant sensing, human take-over, staged rollout starting over non-populated land.

Getting started

1. Pick one asset class or route with high access cost and low population density; capture the baseline cost.
2. Talk to your national aviation authority early; do a SORA-based risk assessment or use a vendor-provided authorisation package.
3. Pilot one docked system (Path A) for 3 months, with remote pilot and measured intervention rates.
4. Connect imagery to a defect-analytics pipeline and a work-order system; keep inspector sign-off.
5. Run a DPIA and define data retention before the first flight over public areas.

Sources

1. EASA — Drones (UAS) regulatory framework, open and specific category — https://www.easa.europa.eu/en/domains/drones-air-mobility/drones-air-mobility-landscape
2. Commission Implementing Regulation (EU) 2019/947 on rules for the operation of unmanned aircraft — https://eur-lex.europa.eu/eli/reg_impl/2019/947/oj
3. Commission Implementing Regulation (EU) 2021/664 on a regulatory framework for the U-space — https://eur-lex.europa.eu/eli/reg_impl/2021/664/oj
4. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
5. JARUS — SORA (Specific Operations Risk Assessment) — http://jarus-rpas.org/publications/
6. PX4 Autopilot documentation — https://docs.px4.io/
7. ArduPilot documentation — https://ardupilot.org/
8. Skydio — Dock for X10 — https://www.skydio.com/
9. Wingcopter — https://wingcopter.com/
10. EDPB — Guidelines 3/2019 on processing of personal data through video devices — https://www.edpb.europa.eu/our-work-tools/our-documents/guidelines/guidelines-32019-processing-personal-data-through-video_en