TECHNICAL DATA
SYSTEM OVERVIEW
01
SUB-01 / SOFTWARE / VISION & GUIDANCE
VISION & AUTONOMOUS GUIDANCE
Burak Selek · İpek Arslan — image processing / Serdar Anıl Demirbaş — integration
Detection runs onboard. A YOLO11m model executes on an NVIDIA Jetson Orin NX 16GB, reading a 1080p RTSP stream from a SIYI A8 mini gimbal camera with an 81° horizontal field of view. The flight configuration is full-frame inference at 640 px, which holds 15–20 FPS on the aircraft — fast enough for the autopilot to act on a detection while the target is still in frame.
The first model failed in the field: detections were lost at altitude. The root cause was a dataset without small-object examples and scale augmentation left disabled. Rebuilding the dataset — 16,015 images merged, deduplicated and class-balanced — and retraining at higher resolution brought mAP50 to 0.950 on a validation set built only from real aerial footage, not web imagery.
Confirmed detections are handed to the autopilot as a position, so the aircraft positions itself over the target autonomously instead of waiting for an operator.
READOUT
MODEL
YOLO11m
ONBOARD COMPUTE
NVIDIA Jetson Orin NX 16GB
SENSOR
SIYI A8 mini · 81° HFOV · 1080p RTSP
FLIGHT CONFIG
640 px full-frame · 15–20 FPS
mAP50
0.950 — aerial-only validation set
DATASET
16,015 images · deduplicated, rebalanced
02
SUB-02 / SOFTWARE / MAPPING
MAPPING — FASTMOSAIC
Ünal Namdar · Simay Akgündüz — mapping specialists
Off-the-shelf photogrammetry was too slow and too opaque for a competition timeline, so the team wrote its own. FastMosaic selects frames using GPS telemetry, then aligns them with ORB feature matching and RANSAC outlier rejection, writing the mosaic straight out as a PNG in submission format — no intermediate point cloud, no third-party desktop stage.
It also runs during the flight. A live coverage preview shows the operator which part of the search area has already been mapped, so gaps are caught while the aircraft is still airborne instead of after landing.
READOUT
PIPELINE
In-house — FastMosaic
FRAME SELECTION
GPS-assisted from telemetry
ALIGNMENT
ORB features + RANSAC
OUTPUT
PNG, submission format, no post-stage
IN-FLIGHT
Live coverage preview for the operator
03
SUB-03 / SOFTWARE / SIMULATION & INTEGRATION
SIMULATION & SYSTEM INTEGRATION
Serdar Anıl Demirbaş — software team leader, system integration
Nothing new flies before it has flown in simulation. The environment combines Gazebo, ArduPilot SITL, ROS 2 and pymavlink — but the part that matters is the architecture. It is distributed: the real Jetson joins the network as its own node and camera imagery arrives over RTSP exactly as it does in flight. The flight software is therefore exercised on flight hardware without the aircraft leaving the bench.
Failure cases are rehearsed here first. GPS degradation and EKF failure, telemetry link loss and low-battery thresholds were all triggered in SITL, and the autopilot response — loiter, altitude hold, continue-then-return, geofence return — was verified before any of it was trusted in the air.
READOUT
STACK
Gazebo + ArduPilot SITL + ROS 2 + pymavlink
ARCHITECTURE
Distributed — real Jetson in the loop
VIDEO PATH
RTSP, identical to the flight path
FAILSAFES REHEARSED
GPS/EKF loss · link loss · low battery · geofence
AUTOPILOT / GCS
ArduPilot · Mission Planner
04
SUB-04 / MECHANICAL TEAM
AIRFRAME & STRUCTURES
Suavi Yiğit Ölmez — lead / Muhammed Arslan · Mehmet Afşın Demirkasımoğlu · Yiğit Karkın · Atakan Akcan · Yaren Civan
The configuration was not chosen by preference. Candidate layouts were scored with the Analytic Hierarchy Process on the Saaty 1–9 scale, turning engineering judgement into a pairwise comparison matrix. The quadrotor won on the combined criteria, and the decision is reproducible rather than asserted.
The frame is deliberately hybrid. Carbon fibre booms carry the load, while the central plates are 2 mm G10 composite. G10 is denser than carbon fibre, and that is the trade: it is RF-transparent, so the avionics deck does not become a Faraday cage around the antennas. Finite element analysis was run on the arm-to-body joints and motor mounts for the 20 kg peak thrust case.
The structure is designed to a 3.33 g limit load factor — aggressive next to the 1.5–2.0 g envelope typical of commercial heavy-lift platforms — which is what buys wind rejection and rapid acceleration rather than headline speed.
READOUT
CONFIGURATION
Quadrotor — selected by AHP, Saaty 1–9
MATERIALS
Carbon fibre booms + 2 mm G10 plates
WHY G10
RF-transparent — protects the comms link
LIMIT LOAD FACTOR
3.33 g
Vne
20 m/s — sized for 15 kt gust loading
ANALYSIS
FEA on arm joints and motor mounts @ 20 kg
OPERATING WEIGHT / MTOW
5.8 kg / 10 kg
05
SUB-05 / ELECTRONICS & COMMUNICATIONS
AVIONICS & COMMUNICATIONS
Zeynep Sena Polat — lead / Utku Oğul Bolat — ground control station operator
The power and propulsion chain was sized before anything was bought. A power budget built from estimated draw set the battery, the DC-DC converters and the connectors. The result is roughly 20 kg of peak static thrust against a 6.5 kg mission weight — a 3.3:1 thrust-to-weight ratio, falling to 2:1 at the 10 kg structural MTOW. That margin exists to reject wind, not to win races.
It was validated on a calibrated static thrust stand before it flew: throttle stepped to 100 % while motor temperature, battery voltage sag and ESC telemetry were logged. Telemetry and MAVLink were verified through Mission Planner, and the combined video and control link was bench-tested for zero packet drop at full throughput.
The onboard computer is a live constraint, not a free addition. Its 25 W peak draw pushed battery capacity up and pushed directly against the weight target — a trade the team took deliberately in order to keep detection onboard.
READOUT
FLIGHT CONTROLLER
Pixhawk-class · ArduPilot firmware
GROUND STATION
Mission Planner
PEAK STATIC THRUST
≈ 20 kg
THRUST-TO-WEIGHT
3.3:1 @ 6.5 kg · 2:1 @ 10 kg MTOW
ENDURANCE
≈ 24 min @ 6.5 kg · 15 min @ 10 kg
COMPUTE POWER BUDGET
25 W peak — Jetson Orin NX
BENCH VALIDATION
Static thrust stand to 100 % throttle
06
SUB-06 / MECHANICAL + AVIONICS — JOINT
PAYLOAD & RELEASE
Mechanical team and electronics & communications team, jointly
The release system is a motorised winch with a servo-actuated lock on a reinforced sub-frame. The lock holds with zero power draw — and that was tested rather than assumed: a payload matching the competition suite in mass was hung from it and left, then checked for slippage over time.
Actuation is commanded through the Jetson control loop, and the bench test measured the delay between the deployment signal and the physical release. Drop reliability was exercised from a range of altitudes before the mechanism was ever flown loaded.
READOUT
MECHANISM
Motorised winch + servo-actuated lock
HOLDING
Zero-power lock, slip-tested under static load
COMMAND PATH
Jetson → autopilot → release
BENCH TEST
Signal-to-release latency measured
FLIGHT VALIDATION
Multi-altitude drop trials before loaded flight
TEAM TO FILL
Measured release latency (ms) and drop accuracy
FIELD EVIDENCE / END-TO-END TEST
ONE COMPLETE RUN, START TO FINISH
In the one comprehensive end-to-end field test the team has flown, the aircraft found both targets — the mannequin and the tent — from roughly 20 m above ground, handed their positions to guidance, flew to each of them autonomously and released the payloads. The water bottle landed about 5 m from the mannequin.
Detection, localisation, guidance and release ran as one chain, with no operator taking over in the middle. That is the whole point of the six subsystems above: individually they are components, and together they either close the loop or they do not. On that run, they closed it.
n = 1
This is one comprehensive test, not a statistic. It is the strongest evidence the team currently holds, and it is reported as exactly that: a single successful end-to-end run. Repeat trials are the next item on the schedule, and this page will be updated with the sample size once they are flown.
0.950
mAP50 · AERIAL-ONLY VALIDATION SET
16,015
IMAGES · CONSOLIDATED DATASET
15–20
FPS ONBOARD · 640 PX FULL-FRAME
~20 m
AGL · BOTH TARGETS DETECTED
METHOD / VERIFICATION LADDER
HOW WE KNOW IT WORKS
SIX HABITS, APPLIED TO EVERY SUBSYSTEM
SIMULATE FIRST
Autonomous mission logic, waypoint execution and failsafe triggers are exercised in Gazebo and ArduPilot SITL before any hardware is at risk. Crosswind, bad weather and sensor failure are cheap in software and expensive in the field.
BENCH BEFORE FLIGHT
Propulsion is run on a calibrated static thrust stand to full throttle with motor temperature, voltage sag and ESC telemetry logged. Sensors are calibrated against EMI, the data link is tested for zero packet drop, and the release lock is load-tested for slippage.
PROGRESSIVE FLIGHT LADDER
Four sequential stages: unloaded manual, unloaded autonomous, loaded manual, loaded autonomous. Aerodynamic stability is confirmed before autonomy is introduced, and autonomy before payload dynamics.
SAFETY PROTOCOL
Props-off policy for indoor bench testing, fireproof LiPo bags, a five-metre safety perimeter before arming, and immediate manual takeover below five metres altitude.
DECIDE WITH NUMBERS
The configuration was chosen with the Analytic Hierarchy Process rather than by preference. The power budget was computed from estimated draw before procurement. Structural joints were checked with FEA at peak thrust.
REVISE FROM TEST DATA
The vision model was rebuilt because it failed in the field, not because it looked wrong on paper. Every flight is followed by inspection for loosening, cracks, vibration and alignment, and the findings go back into CAD.
