Development and Testing for Perception Based Autonomous Landing of a Long-Range QuadPlane
A YOLO–TensorRT–Isaac ROS QuadPlane landing stack achieved 32.7 ms per frame, enabling over 30 FPS edge inference.
Key Findings
Methodology
The authors convert a Voluntex Ranger 2400 fixed-wing aircraft into a 2.4 m, approximately 4.56 kg QuadPlane. The platform integrates Pixhawk/ArduPilot avionics, an Arducam IMX519 RGB camera, Intel RealSense D435i RGB-D sensing, an FPV camera, and an NVIDIA Jetson Orin Nano Super. A YOLO detector is exported from PyTorch to ONNX and compiled as an FP16 TensorRT engine. Pose estimation uses NVIDIA Isaac ROS Visual SLAM, while nvblox builds a voxel map. Training is upgraded from CARLA UE4 to UE5.
Key Results
- On 120 unseen helipad images collected under diverse outdoor conditions, preprocessing, FP16 GPU inference, and postprocessing required 8.5, 19.1, and 5.1 ms, respectively. The 32.7 ms total frame time corresponds to more than 30 FPS. Detection accuracy, recall, and mAP were not reported, so the paper establishes latency rather than accuracy superiority.
- Prototype I, tested without the Jetson and perception payload, climbed to 15 m, held altitude for 30 s, and landed. The reported roll and pitch angular error relative to desired attitudes remained within 1.5°, providing an initial flight-stability baseline.
- Prototype II incorporates the complete payload, but full-envelope flight, airborne D435i validation, GPS-denied descent, and closed-loop autonomous landing remain future work. The evidence therefore supports system integration and real-time inference, not yet complete landing performance.
Significance
The work addresses a persistent sim-to-real and scale gap in autonomous VTOL landing. Large QuadPlanes have higher inertia, weaker hover authority, and slower responses than small multirotors, amplifying the consequences of perception latency and pose error. By integrating visual detection, RGB-D VIO, embedded computing, avionics, power distribution, and structural reinforcement on one aircraft, the study offers a reproducible engineering foundation for GPS-denied monitoring and urban operations. It also makes clear which validation steps remain before reliable autonomy can be claimed.
Technical Contribution
The principal contribution is a jointly engineered deployment stack rather than a new learning theory. CARLA UE5 supplies more realistic illumination, shadows, and geometry; TensorRT FP16 reduces YOLO inference cost; Isaac ROS Visual SLAM and nvblox share the Jetson with the detector; and carbon-fiber beams, reinforcement plates, and FDM parts accommodate the payload. This co-design treats latency, memory, power, center of gravity, rotor-arm loads, and control stability as one deployment problem, enabling a practical path toward onboard perception on a large QuadPlane.
Novelty
The authors position this as the first integrated hardware–software framework for onboard DNN-based QuadPlane landing in unstructured, GPS-denied environments. Its novelty lies in combining UE5 iterative training, a 2.4 m long-range platform, RGB/RGB-D sensing, VIO, voxel mapping, and a single Jetson Orin Nano Super under real payload constraints. Relative to marker-based landing, small-multirotor demonstrations, and simulation-only studies, it emphasizes deployability. However, the claimed distinction is architectural and engineering; complete autonomous-landing evidence is still pending.
Limitations
- The successful flight test used Prototype I without the Jetson and perception modules. Consequently, it does not demonstrate that Prototype II preserves the same stability, response, power margin, or center-of-gravity characteristics under the full payload.
- The paper reports timing on 120 images but omits detector mAP, recall, YOLO version, training-set scale, VIO drift, power consumption, and autonomous-landing success rate, preventing rigorous comparison with state of the art.
- The D435i was functionally checked on the ground only; vibration, ground effect, rapid descent, shadows, reflections, occlusion, and low-texture surfaces remain unvalidated.
Future Work
The authors plan full-envelope testing and aerodynamic numerical analysis of Prototype II, followed by in-flight validation of the D435i, Isaac ROS Visual SLAM, and nvblox. The final experiment is a GPS-denied autonomous sequence covering helipad detection, onboard pose estimation, and controlled descent. Stronger follow-up work should additionally publish accuracy, drift, power, thermal behavior, failure cases, closed-loop controller parameters, and statistically repeated landing trials.
AI Executive Summary
Long-range QuadPlanes combine fixed-wing efficiency with multicopter vertical takeoff and landing, but landing without GPS is unusually demanding. Real sites are unstructured, while large airframes have high inertia, limited thrust vectoring, and slow responses. Existing work often relies on markers, small multirotors, external computing, or simulation, leaving a substantial gap between an impressive virtual demonstration and a deployable aircraft.
This study converts a Voluntex Ranger 2400 into a 2.4 m QuadPlane weighing about 4.56 kg. Its stack includes Pixhawk/ArduPilot avionics, an Arducam IMX519 downward RGB camera, an Intel RealSense D435i RGB-D camera, an FPV camera, and a Jetson Orin Nano Super. The authors upgrade the iterative training pipeline from CARLA UE4 to UE5, use YOLO for helipad detection, compile the model through ONNX and TensorRT FP16, and run Isaac ROS Visual SLAM with nvblox for six-degree-of-freedom pose and voxel mapping. Carbon-fiber reinforcement and FDM-printed mounts address structural and packaging constraints.
On 120 unseen helipad images, preprocessing, inference, and postprocessing took 8.5, 19.1, and 5.1 ms, for 32.7 ms per frame and over 30 FPS. Prototype I climbed to 15 m, held for 30 s, and maintained roll/pitch error within 1.5°. Yet complete payload flight, VIO in the air, detector accuracy, and GPS-denied autonomous landing are not reported. The paper’s strongest result is therefore a credible deployment foundation, not a completed performance validation.
Deep Analysis
Background
Autonomous vertical landing couples perception, state estimation, control, wind disturbance, ground effect, and limited thrust. Early systems used visual markers and Kalman filtering; Falanga et al. demonstrated deep visual landing on a moving platform, while later work explored markerless guidance and RGB-D landing. CARLA, AirSim, and GUAM enabled simulation studies, but many demonstrations target small multirotors, structured zones, UE4 realism, or offboard computation rather than large long-range QuadPlanes.
Core Problem
The target problem is real-time detection of a usable landing site, six-degree-of-freedom localization, and stable descent in GPS-denied, unstructured environments. Bottlenecks include Jetson memory and latency, detector–VIO resource contention, limited fuselage volume, rotor-arm loads, slow large-airframe dynamics, and domain shifts in lighting, weather, texture, and clutter. These factors make perception errors especially costly during the final meters of flight.
Innovation
- �� CARLA UE4 is upgraded to UE5 to improve global illumination, shadows, geometry, and simulation diversity.
- �� A 2.4 m, 4.56 kg fixed-wing platform is structurally converted into a sensor-bearing QuadPlane.
- �� YOLO, ONNX, TensorRT FP16, RGB-D sensing, Isaac ROS Visual SLAM, and nvblox are co-deployed on one Jetson.
- �� Carbon-fiber beams, 1.5 mm reinforcement plates, TAROT FY 690S mounts, and FDM-printed inserts address payload integration and wing stiffness.
- �� The study connects algorithm latency with aerodynamic balance, power distribution, and flight validation rather than treating perception independently.
Methodology
- �� Platform input: Voluntex Ranger 2400, 6S LiPo battery, four Flash Hobby D4215 motors, and 12-inch propellers with 5.5-inch pitch; each motor provides about 2 kg thrust, for approximately 8 kg total.
- �� Training input: CARLA UE5 scenes with varied weather, illumination, and backgrounds; iterative YOLO retraining targets helipad generalization.
- �� Inference: a PyTorch model is exported by Ultralytics to ONNX and compiled into an FP16 TensorRT engine; output is a detected landing-pad location.
- �� State estimation: synchronized D435i RGB-D and IMU streams enter Isaac ROS Visual SLAM, producing GNSS-independent 6-DoF pose; nvblox incrementally builds a voxel map.
- �� Validation: Prototype I tests baseline loitering, while Prototype II carries the complete perception payload.
Experiments
The evaluation contains software timing and preliminary flight testing. Software uses 120 unseen helipad images gathered under diverse outdoor conditions and measures preprocessing, FP16 GPU inference, and postprocessing time. No detector accuracy baseline, YOLO version, training-set size, or formal ablation table is given. Flight testing uses Prototype I: the aircraft climbs to 15 m, holds altitude for 30 s, and lands, while roll and pitch error are evaluated. Prototype II integration is complete, but full flight and GPS-denied autonomous landing are future experiments.
Results
The strongest quantitative result is real-time throughput: 8.5 + 19.1 + 5.1 = 32.7 ms per frame, corresponding to above 30 FPS. The baseline aircraft held 15 m for 30 s with roll/pitch angular error within 1.5°. These results demonstrate computational feasibility and basic flight stability, but not landing success. mAP, recall, VIO drift, end-to-end control error, power consumption, and cross-weather performance are absent, so superiority over competing detectors cannot be established.
Applications
Potential uses include long-range aerial monitoring, infrastructure inspection, urban surveying, disaster response, and operations where GNSS is unavailable or unreliable. Deployment requires adequate visual texture and illumination, balanced payload placement, validated ArduPilot tuning, sufficient power, and safe fallback procedures. RGB detection can locate a landing region, while RGB-D, VIO, and nvblox provide local pose and spatial awareness for onboard decision-making under communication constraints.
Limitations & Outlook
Closed-loop autonomous landing has not yet been demonstrated. Prototype I excludes the Jetson, cameras, and D435i, so the effect of the full payload on center of gravity, vibration, power margin, control response, and aerodynamic loads is unknown. A 120-image timing test cannot establish robustness to glare, shadows, occlusion, low texture, weather, or unfamiliar sites. Future work should report detector accuracy, VIO drift, compute and thermal profiles, failure modes, statistically repeated landings, and aerodynamic analyses.
Plain Language Accessible to non-experts
Imagine a delivery vehicle that can travel efficiently on a highway but can also rise and land vertically in a backyard. Its fixed-wing body is the highway system: efficient for long distances. Four rotors are its elevators, allowing operation without a runway. The difficulty is finding a safe parking spot in an unfamiliar yard without using a phone map, then stopping accurately despite the vehicle’s weight and slow steering.
The researchers give the vehicle several kinds of eyes. A fast RGB camera searches for the landing pad. A depth camera estimates how far the ground is, while an FPV camera lets a human monitor the flight. The Jetson is the onboard dispatcher: it turns images into instructions such as “the pad is left; move right.” One program recognizes the pad, while another combines images with motion sensors to estimate where the aircraft is and gradually builds a three-dimensional picture of its surroundings.
Because the new equipment adds weight and stress, carbon-fiber supports and 3D-printed mounts reinforce the aircraft. The computer processed each image in about 32.7 milliseconds, faster than 30 images per second. A simpler prototype climbed to 15 m, stayed there for 30 s, and kept attitude error below 1.5°. The complete aircraft has not yet performed autonomous landing, so this paper is best understood as building and testing the foundation before the final road test.
ELI14 Explained like you're 14
Picture a plane in a video game that has to fly far away, return home without GPS, and land on a random patch of ground. A normal plane can travel efficiently but needs a runway. A drone can hover and land vertically but may not be great at long trips. A QuadPlane combines both abilities—but that also makes it heavier and slower to control. Tricky, right?
The researchers add cameras and a small computer. The RGB camera looks down and searches for the landing pad. The RealSense camera also measures distance, and the FPV camera helps a pilot watch. The Jetson Orin Nano Super is the onboard brain. YOLO spots the pad, while visual-inertial software combines camera views with motion sensors to estimate the aircraft’s position. Another tool builds a 3D map, a bit like your game character mapping a new level while exploring it.
Before flying, the team trains the detector in CARLA UE5, a realistic virtual world with different lights and weather. Then they shrink and speed up the model so the small computer can run it. On 120 new helipad pictures, the whole image-processing cycle took about 32.7 milliseconds—over 30 frames each second. A basic prototype reached 15 m, hovered for 30 s, and kept its tilt error within 1.5°.
But hold on—the complete smart aircraft has not yet completed an autonomous landing! Cameras may behave differently during vibration, shadows, glare, or fast descent. So this is like finishing the engine, sensors, and practice level before entering the championship match. The next challenge is proving that the whole system can land safely on its own.
Glossary
YOLO
YOLO, or You Only Look Once, is a one-stage object detector that predicts object categories and locations from an image in a largely single forward pass. Its design favors low latency for real-time perception.
The paper uses a YOLO-based detector to identify helipads.
TensorRT FP16
TensorRT is NVIDIA’s inference-optimization framework, while FP16 denotes 16-bit floating-point computation. FP16 can reduce memory use and increase throughput when numerical accuracy remains acceptable.
The PyTorch model is exported to ONNX and compiled into an FP16 TensorRT engine.
Visual-Inertial Odometry
VIO estimates motion by fusing camera observations with inertial measurements from an IMU. It can operate without GNSS but is vulnerable to weak visual texture, rapid motion, and accumulated drift.
Isaac ROS Visual SLAM processes D435i RGB-D and IMU data for 6-DoF pose.
nvblox
nvblox is a GPU-accelerated incremental mapping system that converts depth observations into voxel-based spatial representations. It supports real-time obstacle and geometry awareness.
The paper combines nvblox with VIO to construct a 3D environmental map.
CARLA UE5
CARLA running on Unreal Engine 5 is a simulation environment for generating visually rich autonomous-vehicle scenarios. Improved lighting, shadows, and geometry can reduce the simulation-to-reality gap.
The authors migrate iterative landing-perception training from CARLA UE4 to UE5.
Open Questions Unanswered questions from this research
- 1 Does the full payload alter aerodynamic stability, control latency, or energy consumption? Prototype II needs full-envelope flight, aerodynamic modeling, and controlled comparisons.
- 2 What is the joint failure probability of YOLO and VIO under glare, shadows, vibration, low texture, occlusion, and rapid descent? Accuracy, drift, and landing success are not reported.
- 3 When detection, VIO, and mapping share one Jetson, how do sustained thermal load, power use, and resource contention affect closed-loop control?
Applications
Immediate Applications
Long-range aerial monitoring
Monitoring agencies could use fixed-wing efficiency for forests, pipelines, or urban infrastructure, then land vertically near points of interest. Required preparation includes payload balancing, lighting tests, ArduPilot tuning, power validation, and safe-abort procedures.
GNSS-denied inspection
In dense urban areas, tunnel approaches, or disaster zones, RGB detection can locate a candidate landing region while D435i, Isaac ROS Visual SLAM, and nvblox support local positioning. The environment must provide sufficient texture and the aircraft must sustain onboard computation.
Long-term Vision
Autonomous aerial service networks
A mature system could autonomously land at temporary sites for inspection, resupply, emergency communications, or environmental sensing. Major barriers include weather, moving obstacles, certification, reliable fail-safe behavior, and statistical proof of end-to-end safety.
Abstract
QuadPlanes combine the range efficiency of fixed-wing aircraft with the maneuverability of multi-rotor platforms for long-range autonomous missions. In GPS-denied or cluttered urban environments, perception-based landing is vital for reliable operation. Unlike structured landing zones, real-world sites are unstructured and highly variable, requiring strong generalization capabilities from the perception system. Deep neural networks (DNNs) provide a scalable solution for learning landing site features across diverse visual and environmental conditions. While perception-driven landing has been shown in simulation, real-world deployment introduces significant challenges. Payload and volume constraints limit high-performance edge AI devices like the NVIDIA Jetson Orin Nano, which are crucial for real-time detection and control. Accurate pose estimation during descent is necessary, especially in the absence of GPS, and relies on dependable visual-inertial odometry. Achieving this with limited edge AI resources requires careful optimization of the entire deployment framework. The flight characteristics of large QuadPlanes further complicate the problem. These aircraft exhibit high inertia, reduced thrust vectoring, and slow response times further complicate stable landing maneuvers. This work presents a lightweight QuadPlane system for efficient vision-based autonomous landing and visual-inertial odometry, specifically developed for long-range QuadPlane operations such as aerial monitoring. It describes the hardware platform, sensor configuration, and embedded computing architecture designed to meet demanding real-time, physical constraints. This establishes a foundation for deploying autonomous landing in dynamic, unstructured, GPS-denied environments.