Maritime Rescue Drone

Working design decisions

Seven crucial decisions backed by the literature, to be revisited when a budget is established.

• Proposal

  Fixed-wing for cruise and loiter. Add VTOL only if launch and recovery sites demand it, because it costs mass, drag and transition complexity.

• Proposal

  Treat solar as a supplement to the battery, not a replacement, at small-UAV scale.

• Proposal

  Pair thermal and RGB imagery, screen onboard, and keep a human confirming every detection to start.

• Proposal

  Layer the links: direct RF near shore, satellite for offshore command and telemetry, mesh only where every hop holds.

• Proposal

  Fuse multi-constellation GNSS with an inertial unit, add independent heading and altitude, and define the lost-navigation response in advance.

• Proposal

  Build hybrid: carbon fiber for loaded structure, 3D-printed interfaces, mounts and enclosures, and coupon-test any printed structural part first.

• Proposal

  Start avionics from PX4 and Pixhawk open designs, and validate the assembled aircraft rather than the bare board.

Solar airframes

Flexible cells have lengthened small-UAV flights, but they supplement a battery. What counts is power delivered per added gram over the whole mission, not panel nameplate efficiency.

• Checked

  The Air Force Research Laboratory fitted flexible cells to the removable wing sections of a Raven hand-launched UAV. Three prototypes showed 60% longer endurance, and the lab said 100% looked achievable. Each 20 cm² cell bonds to the wing under clear film, and a power module with maximum-power-point tracking feeds the battery.1

• Checked

  Maryland's Robo Raven III, a flapping-wing micro air vehicle, covered under half its wing with cells that produced about 3.6 W in full sun at roughly 6% efficiency. Its motors drew about 30 W, an order of magnitude more than the solar output.2

• Checked

  NASA's Centurion, a 206-foot solar flying wing, was designed with arrays over about 80% of the upper wing producing up to 31 kW at summer noon. Its first flights ran on batteries. It shows what large solar aircraft look like, not what a small drone can expect.3

• Reported

  One European roll-to-roll CIGS project reached 20% cell efficiency in the lab, but its pilot-line modules came in around 11.7%. Compare installed module output, not cell records.4

• Open

  No source found establishes a general efficiency penalty for wing curvature. The effect depends on the cell stack, the wing's actual curvature, and how sunlight meets each part of it.

• Proposal

  Lay out cell boundaries from the real wing surface. Flexible sheets follow gentle one-direction curvature but tend to wrinkle on compound curvature. Series strings lose output when part of the array is shaded or faces a different way in a bank, so string layout, bypass protection and power-point tracking matter as much as the cells.

Long-range links

No single long-range radio exists. Plan layers, and treat every advertised range as a best case.

• Checked

  ArduPilot documents open-source SiK telemetry radios at up to 100 mW, with typical range near 500 m on standard antennas. One test flight held its link to nearly 3 km with a ground patch antenna.1

• Vendor claim

  Higher-power SiK variants advertise much longer line-of-sight ranges. Those are best-case product figures, not tested operating ranges.2

• Checked

  Radio horizon is roughly 4.1 × (√h₁ + √h₂) km, with antenna heights in metres. A 2 m ship antenna and a drone at 120 m give about 51 km in ideal geometry. Sea reflections, low antennas and link margin make usable range shorter.

• Reported

  Sea-surface reflections can add to or cancel the direct signal, so antenna height, polarization and first-Fresnel-zone clearance matter as much as transmit power.3

• Checked

  Satellite integrator Ground Control lists Iridium Certus 100 at 22 kbps up and 88 kbps down. That suits command, telemetry and stills, not live video.4

• Vendor claim

  Mesh radio makers advertise ranges of hundreds of kilometres. A mesh only extends reach if every hop has its own viable link, and relays add latency and use airtime.5

• Reported

  ADS-B (978 MHz UAT or 1090 MHz Extended Squitter in the US) broadcasts position for air-traffic awareness. It is not a control link.6

• Reported

  A US FCC rule effective February 7, 2025 set up direct frequency assignments in part of the 5030–5091 MHz band for UAS control links. It does not make any 5 GHz radio legal to fly.7

Thermal imaging and onboard AI

Thermal drone payloads are mature. Autonomous person detection over open water is not independently validated, so design is set for AI triage with a human deciding.

• Checked

  SeaDronesSee, the main public UAV benchmark for people in open water, has over 54,000 annotated frames and 400,000 labeled instances, captured from 5 to 260 m altitude at viewing angles of 0 to 90°.1

• Reported

  Its large dataset is visible-light imagery, so thermal detection needs thermal training data of its own.1

• Checked

  A 2024 Sensors paper trained an improved YOLOv5s on 5,736 thermal images of aquatic scenes plus synthetic images, reporting 85.49% average detection accuracy at 23.51 FPS. That is the authors' own dataset, not a drone-at-sea test, and the authors flag tiny targets and seasonal and weather variation.2

• Reported

  A 2025 survey of maritime SAR imagery finds people are tiny relative to the search area, waves and foam cause false alarms, and labeled real-world data is scarce.3

• Reported

  Element 84's SAR pipeline screens video with a fast thermal model, passes candidates to a detailed optical detector, and geolocates hits from flight metadata and camera calibration. It ran on a Steam Deck, not a Jetson.4

• Vendor claim

  NVIDIA lists Jetson Orin NX at up to 100 TOPS in a configurable 10–25 W envelope. TOPS do not predict frame rate for a given detector.5

• Vendor claim

  DJI's search-and-rescue pages describe thermal payloads of 640×512 (Matrice 30T) and 1280×1024 (Zenmuse H30T) pixels, with stated flight times of 41 and 59 minutes. The workflow described is operator-driven rather than validated automatic detection over water.6

• Open

  No public, independently evaluated benchmark was found for thermal AI detecting real people in open water from a drone, across sea states, weather and day or night.

Navigation

Better GPS isn't a navigation solution. Combine satellite fixes with inertial sensing and independent checks, and decide what the aircraft does when they disagree.

• Reported

  GPS, GLONASS, Galileo and BeiDou are independent constellations. Tracking several improves availability and satellite geometry, but does not protect against jamming, spoofing or receiver faults.1

• Reported

  RTK delivers corrections to the aircraft during flight. PPK refines the recorded track afterward, which helps survey data but cannot serve as live navigation.2

• Reported

  PX4's estimator fuses inertial data with GNSS and other sensors to estimate position, velocity, orientation and sensor biases. An inertial unit rides through short GNSS outages, but its error grows with time.3

• Reported

  Dual-antenna GNSS/INS units such as VectorNav's VN-300 give heading without a magnetometer or vehicle motion.4

• Checked

  Galileo's OSNMA, which lets a receiver verify that navigation data came from Galileo unaltered, was declared operational on July 24, 2025. It authenticates data; it does not stop jamming.5

• Reported

  ArduPilot's dead-reckoning failsafe documentation notes that many vehicles can keep control without GPS for only about 10 to 30 seconds.6

• Reported

  AIS shows vessels that carry transponders. It is one awareness input, not detect-and-avoid.7

Fabrication

Print the interfaces, not the spars. FDM for iteration, SLS nylon for tougher small batches, and metal only for a justified part with a partner lab.

• Reported

  FDM is cheapest and fastest for iteration but weakest between layers. SLS nylon suits durable brackets and housings. SLA gives fine detail, but many resins are brittle and degrade in UV. Metal additive manufacturing needs specialist equipment and inspection.1

• Reported

  In a 2023 study of multirotor airframe specimens, FDM and SLS parts reached up to about 40 MPa, while continuous-fiber specimens reached at least four times that. Print direction changed the results.2

• Reported

  A published multirotor design framework printed custom joints and mounts and combined them with carbon-fiber tubes and plates, the hybrid approach proposed here.3

• Reported

  A 2024 flying-wing prototype printed in PLA on a desktop printer was a proof of concept. Its authors say durability needs more investigation.4

• Reported

  A 2025 test-stand study of printed propellers found resin gave the highest thrust and PETG the best resistance to axial loading. Print propellers only for contained test-stand work.5

• Reported

  General Atomics and Norsk Titanium described a printed Ti-6Al-4V wing splice for an unmanned aircraft. The article does not show flight qualification.6

Airframe and endurance

Choose the configuration from the mission: fixed-wing for distance and loiter, multirotor only where hovering is the job.

• Reported

  Maximum endurance and maximum range occur at different airspeeds, and both shift with weight and altitude. Optimize at the speed the aircraft will actually fly.1

• Checked

  Induced drag falls as aspect ratio rises (C_Di ≈ C_L² / (π·e·AR)), so long slender wings suit lift-heavy missions.

• Reported

  High-aspect-ratio wings can be soft in bending and torsion, which lowers flutter speed, so aeroelastic stability belongs in the early design loop.2

• Reported

  For multirotors, endurance and optimal speed come from the whole propeller, motor and battery system, not from hover power alone.3

• Reported

  A NASA study of high-altitude long-endurance aircraft found that maximizing endurance alone can give a poor mission solution, with energy-storage density and solar-cell efficiency as key drivers.4

• Reported

  Morphing wings can in principle adapt to climb, cruise and loiter, but actuators, skins and controls add mass and complexity. Benefits must be judged for the whole aircraft.5

• Reported

  The Sun Falcon 2 paper documents a design loop linking aerodynamics, airframe mass, battery and solar energy. It reports only short initial flight tests, not its multi-day goal.6

• Reported

  ETH Zurich's AtlantikSolar is a practical research example of a solar UAV built for multi-day flight.7

Avionics and PCB design

Keep high-current power, quiet sensors and RF apart on the board, then prove the arrangement works in the assembled aircraft.

• Reported

  Pixhawk open standards and PX4 reference designs document controller architecture, connectors, pinouts and layout guidance, a sound starting point for a compatible board.12

• Reported

  A power distribution board feeds the ESCs and motors. A power module also supplies regulated flight-controller power and the battery measurements used for failsafes.3

• Reported

  IPC-2152 relates conductor size, copper weight and board construction to current and temperature rise. Validate critical high-current paths with thermal measurement, not only a trace-width calculator.4

• Reported

  PX4 warns that vibration can clip sensors and break the estimator. Fix it at the source, with balanced propellers and quality parts, before relying on software filters.5

• Reported

  Mount the controller near the center of gravity in a known orientation, or configure IMU position offsets.6

• Proposal

  Likely noise sources are ESC switching, motor wiring, DC-DC converters and radios. Keep sensitive circuits, antennas and magnetometers away from switching loops and high-current paths, and test GNSS and links with the motors running.

Fleet coordination

Commercial drone swarms are mostly human-supervised fleet platforms. Autonomous cooperative search is mostly research.

• Vendor claim

  Skydio and DJI publicly describe alert-to-dispatch workflows: an alert supplies coordinates, software picks a ready aircraft or dock, an operator confirms, video is shared and records are archived. Neither publishes a full backend design.12

• Reported

  Multi-UAV search research splits a region into cells, allocates them across aircraft with different payloads and endurance, and routes around no-fly zones. One platform was demonstrated in a real SAR scenario.3

• Reported

  Dynamic task-assignment research clusters tasks, assigns them locally, and uses a central controller to relocate idle aircraft. Static plans fail when tasks appear or aircraft drop out.4

• Reported

  Fleet software is not airspace management. NASA's UTM concept makes separation a shared responsibility among service suppliers, aircraft and other airspace users.5

• Proposal

  A hybrid pattern suits rescue: a central service for assignment, policy and audit, and onboard logic for time-critical local decisions.

• Proposal

  Keep incident, mission and flight as separate records, since one call can start several missions and one mission several flights. Store event history rather than only current state, telemetry in a time-series store, and video in object storage with integrity hashes.

Incident ──► Mission ──► Flight ──┬──► Telemetry   time series
                                  ├──► Media       hashed, object store
                                  └──► Detections  confidence + reviewer
Audit log ──► every view, command and export

Ohio State resources

Most of what is useful is gated by a faculty sponsor, training or project approval. Most resources require approval or an application.

What is still unproven

Where the evidence is thin, so nobody mistakes a plan for a result.

• Open

  Thermal-AI person detection from drones in real sea states has no independent benchmark. Closing it takes thermal footage of people in water under varied conditions, scored per condition.

• Open

  The benefit or penalty of curved solar cells on a UAV wing is design-specific. Only an installed-article test in flight settles it.

• Open

  Mesh relay range figures are almost all vendor claims. Test hop by hop at the real altitudes.

• Open

  Printed structural parts show encouraging results but need orientation-aware, fatigue and vibration testing, and printed propellers belong only on guarded stands.

• Open

  Commercial fleet backends are not public, so the data model in these notes is a proposal.

• Open

  The beyond-visual-line-of-sight approval pathway is jurisdiction-specific, and the FAA pages I needed were not retrievable. Check current waiver and authorization terms for the real route.