Funke Fully Autonomous Flight Challenge

• Funke Fully Autonomous Flight Challenge is a rigorous test of end-to-end UAV autonomy in GNSS-denied environments, integrating perception, mapping, localization, planning, and control.
• It mandates 9 km waypoint navigation below 25 m AGL without GNSS or dense pre-mapping, driving innovations in drift correction and real-time obstacle avoidance.
• Implementations like the winning SPRIN-D system demonstrate effective CPU-only onboard processing and highlight trade-offs between accuracy, compute limitations, and hardware design.

The Funke Fully Autonomous Flight Challenge is a GNSS-denied UAV autonomy problem centered on fully onboard navigation over long distances and under stringent operational constraints. In the winning SPRIN-D deployment, the mission required 9 km long-range waypoint navigation below 25 m AGL (Above Ground Level) without GNSS or prior dense mapping, making the challenge a coupled problem in perception, mapping, localization, planning, and control rather than a single-module benchmark (Werner et al., 1 Oct 2025). Multiple later and contemporaneous works explicitly map their methods to this challenge or to the same requirement profile, including unknown-environment navigation, real-time obstacle avoidance, embedded execution, and robustness under severe sensing and compute limits (Lu et al., 2024, Guan et al., 12 Jan 2026, Gu et al., 6 Feb 2026, Metwally et al., 17 Feb 2026).

1. Challenge definition and task structure

A concise operational characterization of the challenge is given by the winning system paper: reliable long-range flight of unmanned aerial vehicles in GNSS-denied environments is difficult because integrating odometry leads to drift, loop closures are unavailable in previously unseen areas, and embedded platforms provide limited computational power (Werner et al., 1 Oct 2025). The same source identifies the formal mission envelope as 9 km long-range waypoint navigation below 25 m AGL without GNSS or prior dense mapping, with waypoint detection performed on flags on 1 m poles (Werner et al., 1 Oct 2025).

Other papers that explicitly frame their contributions against the Funke challenge broaden that characterization. One summary states that the challenge calls for fully onboard perception and mapping in unknown, cluttered 3D environments, high-speed, aggressive obstacle avoidance, real-time guaranteed-feasible trajectory planning under full vehicle dynamics, precise tracking of planned maneuvers up to platform limits, and robustness to unmodeled disturbances (Lu et al., 2024). A separate architecture paper formulates the same problem in operational-autonomy terms as full autonomy: no manual intervention once mission starts, together with multi-mode flight, real-time obstacle detection/avoidance and mission replanning, and high reliability and safety under embedded constraints (Metwally et al., 17 Feb 2026).

Taken together, these descriptions show that the challenge is not reducible to waypoint following. It is a systems problem in which localization observability, planner feasibility, actuator limits, perception latency, and embedded software determinism all directly affect mission completion. This suggests that Funke functions less as a narrowly defined navigation contest than as a stress test of end-to-end aerial autonomy.

2. Reference implementation: the winning SPRIN-D system

The most concrete published realization is the fully onboard system that won the SPRIN-D Funke Fully Autonomous Flight Challenge (Werner et al., 1 Oct 2025). Its autonomy stack runs entirely on an Intel NUC i7 (16 GB RAM) without GPU. The perception suite combines a Livox Mid-360 LiDAR producing point clouds at ~100 kpts/s, an Intel RealSense D435 at 30 Hz, a Bluefox2 global-shutter RGB + ICM42688 IMU for monocular VIO via OpenVINS, and a magnetometer for global yaw alignment (Werner et al., 1 Oct 2025).

Its mapping and planning pipeline is explicitly layered. LiDAR populates an OctoMap local occupancy map (40 × 40 m, updated at 10 Hz); occupancy is converted to an Euclidean Signed Distance Field and then to an A* graph; the resulting path is passed to a polynomial trajectory generator and then to Model Predictive Reference Tracking (MRS system) for motor commands (Werner et al., 1 Oct 2025). The localization module is the distinctive element: a local heightmap (1 m bins) is extracted from occupancy, converted to finite-difference gradients, thresholded into a binary edge map, template-matched against a prior DEM-derived edge map, and fused with odometry in a clustered particle filter (Werner et al., 1 Oct 2025).

The localization state is deliberately low-dimensional. Each particle stores only a 2D position $x_t^i=(x^i,y^i)^\top$, while yaw is taken directly from the compass. Prediction uses odometry displacement $u_t=(\Delta x,\Delta y)$ plus Gaussian process noise, measurement weights are derived from normalized template-matching scores through

$w_t^i \propto \exp\!\bigl(\kappa\,(z_t^i-1)\bigr),$

with $\kappa \approx 10$, and multimodality is handled by K-means with $K=3$ followed by selecting the centroid of the largest cluster as the corrected $(x,y)$ (Werner et al., 1 Oct 2025). The resulting pipeline runs in real time: OctoMap, ESDF, and A* each run at 10 Hz; heightmap build + gradient + template-matching run at ~1 Hz; the particle filter (1,000 particles) also runs at 1 Hz; and total localization CPU load is reported as < 15 % (Werner et al., 1 Oct 2025).

Field results establish the challenge’s practical scale. The system flew eight GNSS-denied missions (495 m–1 371 m). For the reported flights, raw odometry RMSE values of 36 m, 31 m, 27 m, and 53 m were reduced to 13 m, 10 m, 6 m, and 7 m respectively; averaged over kilometer-scale legs, drift was reduced by ~60 %, from ~32 m to ~12 m (Werner et al., 1 Oct 2025). In terrain with strong structural gradients, such as urban and forest regions, errors were reduced to 5–10 m, whereas in open fields the filter gracefully fell back on odometry (Werner et al., 1 Oct 2025). Waypoint detection performance was 85 % within a 15 m search radius using YOLOv8n on CPU (Werner et al., 1 Oct 2025).

3. Localization and perception strategies in challenge-oriented research

The challenge-oriented literature does not converge on a single perception doctrine. Instead, it spans geodata matching, VIO, monocular regression, and body-frame perceptual abstractions.

The winning SPRIN-D system uses heightmap gradients as a coarse but robust drift-correction cue, explicitly recommending coarse but robust features (height gradients) rather than full imagery models for real-time on-board drift correction (Werner et al., 1 Oct 2025). By contrast, “Deep Drone Racing” formulates perception as direct regression from image to waypoint + desired speed, where the network outputs normalized image coordinates $[x,y]$ and a normalized speed $v$, and these are back-projected into the vehicle body frame for minimum-jerk planning (Kaufmann et al., 2018). That paper emphasizes that body-frame outputs remove dependence on drift-prone SLAM, a property directly relevant to GNSS-denied challenge settings (Kaufmann et al., 2018).

A more classical modular approach is represented by “Fast, Autonomous Flight in GPS-Denied and Cluttered Environments”, which combines SVO visual-inertial odometry, UKF sensor fusion, local 3D occupancy mapping, hybrid A* search, and convex-corridor trajectory optimization (Mohta et al., 2017). That system fuses cameras at 40 Hz, LiDAR-Lite at 20 Hz, and IMU at 200 Hz, with the UKF running at 200 Hz and the planner at ∼ 4 Hz (Mohta et al., 2017). It demonstrates that a challenge-relevant autonomy stack can be built from tightly synchronized estimation and planning modules rather than from end-to-end policy learning.

At the small-scale end of the platform spectrum, Aerobat shows how the same problem becomes constrained by payload and memory. Its onboard compute is a Raspberry Pi Zero 2 W with 512 MB shared DRAM, leaving ~300 MB for algorithms after a bare-bones ROS Noetic installation and sensor drivers, which the thesis identifies as a major limitation for off-the-shelf VIO libraries (Salagame, 2023). The implemented onboard estimator is only a complementary filter at ~150 Hz, while full 6-DOF EKF/UKF and visual-inertial odometry remain future work; camera–IMU extrinsics and time offset are reported as an open issue (Salagame, 2023). This is important because it illustrates a recurring Funke constraint: autonomy may be limited less by algorithmic availability than by platform mass, power, and memory budgets.

Monocular end-to-end control appears in “Multi-Task Regression-based Learning for Autonomous UAV Flight Control within Unstructured Outdoor Environments”, where a single RGB image is mapped to $\Delta N,\Delta E,\Delta D$ and a desired quaternion (Maciel-Pearson et al., 2019). That method is presented as operating under the forest canopy, regardless of the presence of trails or additional sensors (i.e. GPS) and as requiring only monocular image + IMU (Maciel-Pearson et al., 2019). However, its evaluation is explicitly only in simulation, which distinguishes it from the field-deployed systems above (Maciel-Pearson et al., 2019).

Finally, the perception/control boundary can be formalized architecturally rather than algorithmically. The hybrid F′ and ROS2 architecture demonstrates 87.19 Hz position estimation with 11.47 ms mean latency and 99.90 % data continuity during a 32.25-minute indoor quadrotor flight test, while keeping flight-critical logic in deterministic F′ components and autonomy pipelines in ROS 2 (Metwally et al., 17 Feb 2026). For the Funke problem class, this matters because perception accuracy alone is insufficient if message freshness, scheduling, and fault handling are not also engineered.

4. Planning, control, and airframe diversity

Challenge-oriented planning and control methods span minimum-jerk replanning, minimum-snap corridor optimization, differential-flatness-based tail-sitter planning, MPCC, NMPC under actuator failure, and lightweight flapping-wing attitude stabilization.

For quadrotor-style agile flight, “Deep Drone Racing” couples image-based waypoint regression to a minimum-jerk trajectory planner recomputed in receding horizon, with CNN inference plus back-projection plus polynomial solve taking ≈ 80 ms total, an onboard update rate of ~10 Hz, and a 500 Hz low-level controller (Kaufmann et al., 2018). The system achieved 100% five-lap success up to 9 m/s in static-track simulation and 100% five-lap success on a 21 m real-world race track, with a best lap of ~6.2 s (Kaufmann et al., 2018). The relevance to Funke lies in the coupling of perception and action under limited onboard compute.

For GPS-denied cluttered navigation, Mohta et al. generate a hybrid A* path and then solve a 7th-order polynomial spline problem minimizing snap inside convex corridors, with the QP solved in ∼ 10–20 ms on an Intel NUC (Mohta et al., 2017). LOONG pushes this paradigm toward time optimality by generating a time-optimal polynomial reference at 100 Hz, accelerating time allocation by imitation learning, and then using time-optimal MPCC with safe flight corridor constraints; the reported planning + ref-gen latency is < 1 ms, MPCC solve is ≈6 ms, and the system reaches 18.1 m/s with 10 consecutive trials and 100% success from diverse start points (Guan et al., 12 Jan 2026).

For airframes with highly nonlinear aerodynamics, “Autonomous Tail-Sitter Flights in Unknown Environments” combines FAST-LIO2, guide-path search, safe flight corridor generation, EFOPT trajectory optimization at 5 Hz, and on-manifold MPC at 100 Hz, with low-level PX4 feedback at 400 Hz (Lu et al., 2024). The system reports speeds up to 15 m/s in indoor laboratories, underground parking lots, and outdoor parks, with median solve times ~10–35 ms and worst-case <100 ms (Lu et al., 2024). This directly addresses the challenge formulation that emphasizes guaranteed-feasible planning under full vehicle dynamics.

Fault-tolerant autonomy extends the challenge beyond nominal flight. “Rotor-Failure-Aware Quadrotors Flight in Unknown Environments” integrates online FDD, NMPC, and rotor-failure-aware planning. The controller runs at 200 Hz using ACADO + qpOASES, with average solve time: 1.9 ms (peak 3.2 ms) on a Jetson Orin NX, and uses FDD latencies of 20–30 ms for motor-stop detection and 150–250 ms for propeller damage detection (Zhou et al., 13 Oct 2025). Real-world results include indoor waypoint navigation with moving obstacles, outdoor forest (~30×20 m²) flight, and continued autonomy under rotor failure (Zhou et al., 13 Oct 2025). For Funke-like missions, this paper reframes robustness as post-fault continuity rather than merely nominal performance.

The challenge has also been used to motivate non-quadrotor flight platforms. Aerobat currently achieves open-loop flapping with IMU-based attitude stabilization, 5–7 m indoor untethered flights at 2 Hz flapping, and 10 m outdoor flights in mild wind (≤1 m/s), but no closed-loop position control has yet been achieved (Salagame, 2023). AirPulse, by contrast, reports fully onboard, closed-loop, untethered flight for a 26 g butterfly-inspired FWMAV using a STAR generator, Madgwick filtering at 100 Hz, RLS mean extraction, and PID control on an ESP32-S3 (Gu et al., 6 Feb 2026). The paper explicitly discusses applicability to the Funke challenge, but also states that additional sensing (e.g. stereo vision, LiDAR) and SLAM algorithms are needed for GPS-denied navigation (Gu et al., 6 Feb 2026). This establishes that airframe innovation alone does not solve the challenge; navigation-grade sensing and estimation remain necessary.

System	Platform/method	Reported result
(Werner et al., 1 Oct 2025)	CPU-only NUC i7, LiDAR/VIO/heightmap matching	Drift reduced by ~60 %, from ~32 m to ~12 m
(Lu et al., 2024)	Tail-sitter, EFOPT + MPC	Real-world speeds up to 15 m/s
(Guan et al., 12 Jan 2026)	LiDAR MAV, learned time allocation + MPCC	Peak speed 18.1 m/s, 10 consecutive trials, 100% success
(Zhou et al., 13 Oct 2025)	Rotor-failure-aware NMPC	Autonomous flight in cluttered rooms and unknown forests after single-rotor failure
(Gu et al., 6 Feb 2026)	26 g tailless FWMAV, STAR + onboard PID	Fully onboard, closed-loop, untethered flight

The diversity of these systems indicates that the Funke challenge is airframe-agnostic at the mission level but strongly airframe-dependent at the control and estimation level.

5. Embedded systems, timing, and hardware–software co-design

A consistent lesson across the literature is that embedded implementation details are first-order determinants of autonomy.

The winning SPRIN-D system is explicitly CPU-only, with no GPU, and keeps the entire localization module below 15 % CPU load (Werner et al., 1 Oct 2025). The architecture paper based on F′ and ROS2 reaches a similar conclusion from a software-engineering angle: the system maintained 15.19 % CPU, 1,244 MB RAM, 23.01 % peak CPU, 1,435 MB peak RAM, and zero stale telemetry messages, while all 15 ground commands executed with 100 % success rate (Metwally et al., 17 Feb 2026). Its design principle is a strict partition: flight-critical logic in NASA F′, perception and autonomy in ROS 2, bridged by Protocol Buffers over UDP (Metwally et al., 17 Feb 2026). For challenge conditions, this is significant because it decouples deterministic control timing from higher-level perception variability.

At the opposite end of the resource spectrum, Aerobat shows the consequences of ultra-lightweight compute. With two ROS nodes (IMU + camera) and basic image processing, CPU utilization is ~30 % idle and spikes to ~80 % under load; a bare-bones ROS Noetic plus sensor drivers use ~200 MB, leaving ~300 MB for algorithms (Salagame, 2023). End-to-end IMU-to-control latency is ~6 ms with jitter ±2 ms, achieved through ROS timers rather than rate-based sleeps and Linux FIFO priorities that place imu_node above the controller (Salagame, 2023). The thesis concludes that sensor drivers and ROS scheduling must be trimmed, making hardware–software co-design an explicit requirement rather than an implementation nicety (Salagame, 2023).

Mechanical integration is equally central. The SPRIN-D winner reports that IMU/VIO vibration isolation was critical, stating that without mechanical decoupling, high-frequency motor vibrations caused VIO divergence within seconds (Werner et al., 1 Oct 2025). The rotor-failure-aware system addresses the same issue through airframe hardware, using four anti-torque plates that reduce post-failure peak spin rates from ≈20 rad/s to ≈12 rad/s, thereby preserving mapping accuracy for LiDAR-inertial odometry (Zhou et al., 13 Oct 2025). In challenge settings, these results imply that estimator quality is inseparable from structural vibration management.

Timing and synchronization recur as bottlenecks. Aerobat reports camera–IMU timestamp jitter (~±5 ms) as a factor that complicates future VIO convergence (Salagame, 2023). The F′/ROS2 system uses timestamp acceptance rules so that F′ only accepts messages whose timestamp > last_processed, dropping old packets to prevent stale-control hazards (Metwally et al., 17 Feb 2026). The winning SPRIN-D system identifies magnetometer drift >30° near steel structures as another integration-level issue, noting that relying solely on the compass is unsafe (Werner et al., 1 Oct 2025). These details underscore a central property of the challenge: embedded autonomy fails as often from synchrony, vibration, and stale data as from deficiencies in nominal planner or estimator accuracy.

6. Limitations, misconceptions, and open research directions

A recurring misconception is that fully autonomous flight in the Funke sense is equivalent either to pure odometry or to a single end-to-end neural policy. The literature explicitly mapped to the challenge is more heterogeneous. It includes gradient-template matching fused with odometry in a clustered particle filter (Werner et al., 1 Oct 2025), body-frame waypoint regression plus minimum-jerk planning (Kaufmann et al., 2018), VIO + UKF + occupancy mapping + convex-corridor optimization (Mohta et al., 2017), time-optimal MPCC with learned time allocation (Guan et al., 12 Jan 2026), and fault-aware NMPC (Zhou et al., 13 Oct 2025). This suggests that the challenge is defined less by algorithmic ideology than by operational constraints.

A second misconception is that “fully onboard” implies high-end accelerators. The winning SPRIN-D system runs without GPU on an Intel NUC i7 (Werner et al., 1 Oct 2025), and the F′/ROS2 validation shows substantial scheduling headroom on an embedded Orange Pi 5 while sustaining 87.19 Hz estimation (Metwally et al., 17 Feb 2026). Conversely, the Aerobat platform demonstrates that even modest autonomy can become marginal on sub-100 g vehicles because the 15 g budget forced Pi Zero 2 W rather than typical Odroid/Jetson, leaving VIO only “in the pipeline” (Salagame, 2023). The relevant distinction is therefore not GPU versus CPU, but whether the sensing, compute, and control loops fit the mission envelope of a given airframe.

Published systems also expose concrete unresolved issues. The winning SPRIN-D system reports that in open fields the localization filter falls back to odometry and that limited battery (≈20 min flight) and local-map size constrained maximum range more than drift (Werner et al., 1 Oct 2025). The tail-sitter work notes a LiDAR FoV blind-spot above the belly and that the coordinated-flight assumption excludes uncoordinated side-slip maneuvers (Lu et al., 2024). Aerobat still lacks closed-loop position control, and camera–IMU extrinsics and time-offset remain an open issue (Salagame, 2023). AirPulse explicitly states that additional sensing and SLAM algorithms are needed for GPS-denied navigation (Gu et al., 6 Feb 2026). The monocular MTRL approach is limited by evaluation only in simulation and by the absence of explicit dynamic-obstacle avoidance (Maciel-Pearson et al., 2019).

The most plausible near-term research direction is therefore not a single “best” autonomy stack, but tighter coupling between robust localization backstops, real-time feasible planning, and embedded software architectures that preserve determinism under field conditions. The challenge evidence favors systems that degrade gracefully: when edge structure is weak, odometry dominates; when perception is delayed, old waypoints are reused and smoothed by the planner; when a rotor fails, both control and planning envelopes are reconfigured online (Werner et al., 1 Oct 2025, Kaufmann et al., 2018, Zhou et al., 13 Oct 2025). In that sense, the Funke Fully Autonomous Flight Challenge has become a unifying reference point for research on long-range GNSS-denied autonomy, aggressive cluttered-flight planning, and robust onboard execution across conventional and bio-inspired aerial robots.