GLIDE: Drone-Guided SAR Escort

• GLIDE is a cooperative aerial–ground search-and-rescue system that uses UAVs for long-horizon guidance and precise victim detection to aid UGV navigation.
• It employs a role-separated design where one UAV detects victims with onboard YOLOv11 and another scouts terrain to inform A* based path planning.
• Empirical results from hardware and simulation experiments show that GLIDE reduces rescue times and improves navigation success in complex, non-convex environments.

Searching arXiv for the primary GLIDE paper and closely related escort/navigation work. Guided Long-horizon Integrated Drone Escort (GLIDE) is a cooperative aerial–ground search-and-rescue framework that pairs two unmanned aerial vehicles with an unmanned ground vehicle to achieve rapid victim localization and obstacle-aware navigation in unknown environments. The name denotes the core dependency structure: the UGV is “escorted” and guided by drones that provide long-horizon information unavailable to local onboard sensing. In the reported implementation, a goal-searching UAV performs real-time onboard victim detection and georeferencing, a terrain-scouting UAV flies ahead of the UGV’s planned route to provide mid-level traversability updates, and the UGV fuses these aerial cues with local sensing to run time-efficient A* planning and continuous replanning. The framework was evaluated in hardware using a GEM e6 golf cart and two Holybro X500 quadrotors, with additional simulation ablations that isolate the planning stack from detection (Farrell et al., 17 Sep 2025).

1. Operational problem and design rationale

GLIDE is formulated for search-and-rescue missions in unknown, cluttered, non-convex environments, including disaster zones, collapsed buildings, forests, and flooded streets. In such settings, the environment is partially known, global obstacle structure may contain U-shaped traps and corridors, and mission timing is critical because minimizing time-to-rescue is central to SAR performance. The motivating failure mode is that a UGV using only local, short-horizon sensing, such as LiDAR in a small window, may enter local minima, become trapped, or incur long detours because it cannot infer the global structure of the scene (Farrell et al., 17 Sep 2025).

The framework is explicitly structured around complementary sensing and actuation asymmetries. UAVs provide rapid coverage and a global perspective, but have limited endurance, payload, and often bandwidth-limited communication. UGVs provide persistence, payload, and the ability to physically reach victims, but are constrained by short-horizon sensing and ground-level occlusion. GLIDE addresses this trade-off by role-separated cooperation: one UAV maximizes information about victims, one UAV maximizes information about terrain and obstacles, and the UGV uses those information products in an aerial-guided planning stack (Farrell et al., 17 Sep 2025).

A common misconception is to treat GLIDE as a generic multi-robot team with loosely shared sensing. In the reported formulation, the distinction is sharper. The empirical claim is not merely that multiple robots help, but that explicit role separation across the UAVs, coupled with terrain scouting and guided planning, improves reach time and navigation safety relative to purely local planning or multi-robot configurations without planning integration. The framework is therefore defined as an information-structured system rather than only a platform composition (Farrell et al., 17 Sep 2025).

2. Role-separated architecture and communication

GLIDE consists of one UGV and two UAVs. The UGV is a GEM e6 golf cart; the UAVs are Holybro X500 quadrotors. A TP-Link Archer AX21 router mounted on the UGV provides the network backhaul. UAVs transmit compact georeferenced messages of less than 1 kB per event over Wi-Fi. If connectivity is lost, detections are queued locally and forwarded when the link is restored, so the architecture is designed to tolerate intermittent communication without requiring raw imagery streaming (Farrell et al., 17 Sep 2025).

The goal-searching UAV is responsible for site survey over high-priority regions and victim nomination. It may execute autonomous waypoint surveys uploaded via QGroundControl within a geofence, or be manually piloted for scenario-dependent exploration. Its perception payload is a nadir-facing OakD-Pro-W RGBD camera, and a Jetson Orin Nano runs a lightweight YOLOv11 detector fine-tuned on the VisDrone aerial dataset. The detector outputs bounding boxes for candidate victims at approximately 30 Hz, and each validated georeferenced victim becomes a goal node in the UGV’s planning graph (Farrell et al., 17 Sep 2025).

The terrain-scouting UAV performs escort-style path validation. It flies ahead of the UGV’s current planned route, tracks the ego-vehicle’s path several steps ahead, and, if the current route does not yet reach all victims, follows a straight-line extension toward the next victim in visiting order. In hardware tests it maintains a lead offset, for example 15 m. In the current implementation, much of the semantic and mapping workload remains on the UGV, so the terrain-scouting UAV functions primarily as a mobile sensor head rather than as a fully autonomous mapper (Farrell et al., 17 Sep 2025).

The UGV is the escorted platform and the only vehicle that must physically reach victims. It receives victim goals from the goal-searching UAV, traversability updates from the terrain-scouting UAV and local LiDAR, constructs and maintains a persistent grid-based roadmap, runs A*-based planning, and executes trajectories with a PID controller under drive-by-wire control. In the system flow chart, UAV-side components include YOLOv11 detection, state retrieval via PyMavLink, georeferencing, and ROS2-based communication, while the UGV side includes an aerial mapping node, a planner node, and a control node that sends commands via CAN bus to PACMOD (Farrell et al., 17 Sep 2025).

3. Perception, georeferencing, and map construction

Victim detection is performed from top-down aerial imagery using YOLOv11 from Ultralytics YOLO, trained on the VisDrone dataset with label remapping to operationally relevant classes. The framework maps people to victims and vehicles or walls to non-traversable obstacles. This semantic remapping is important because the same detection backbone contributes to both goal nomination and obstacle insertion, albeit through different UAV roles (Farrell et al., 17 Sep 2025).

Georeferencing is based on camera intrinsics, UAV GPS, altitude, and attitude. For a detection with image coordinates $(u,v)$, the camera ray is first computed by

$\mathbf{r}_c = K^{-1}\begin{bmatrix}u\v\1\end{bmatrix},$

then rotated into the world frame with

$\mathbf{r}_w = R(q)\mathbf{r}_c,$

and intersected with the ground plane under a flat-terrain assumption. The paper reports a closed-form projection from image to world using quaternion and altitude, achieving meter-level accuracy under stable attitudes. RTK corrections improve localization to sub-meter accuracy, while GPS dominates the error budget (Farrell et al., 17 Sep 2025).

The implementation uses several filtering heuristics to control projection error. Detections are discarded when roll or pitch exceeds $15^\circ$. In hardware experiments, flights are conducted at 15 m altitude, only detections near the image center are accepted, and victim insertion requires 3-hit spatial consensus within 1 m. These measures are not presented as a probabilistic state estimator; they are explicit robustness heuristics for a fixed-camera platform without a gimbal (Farrell et al., 17 Sep 2025).

A second misconception is to treat the map as a dense, semantically graded world model. In the planner, traversability is primarily represented as an occupancy-style grid or mid-level map, and non-traversable cells are treated as hard obstacles. In simulation and hardware, GLIDE uses a discrete grid of $300 \times 300$ cells at $0.5$ m resolution in ENU coordinates. The UGV’s aerial mapping node fuses UAV-derived obstacles, UAV footprints, local LiDAR, and RTK pose into a smoothed map over time. Previously unknown regions are not discarded after observation; the map is persistent and accumulates learned structure across the mission (Farrell et al., 17 Sep 2025).

4. Long-horizon planning, proxy goals, and continuous replanning

The central planning mechanism is a grid-based A* planner operating on a uniform grid centered on the UGV. For hardware experiments, the grid corresponds to an approximately $150 \text{ m} \times 150 \text{ m}$ area. Victim coordinates are inserted as goal nodes. If a victim lies outside the current grid bounds, the system creates a proxy goal at the grid boundary along the victim direction, allowing long-horizon progress even under finite map support. When multiple victims exist, the planner computes a pairwise shortest-time or distance matrix across goals and determines a visitation order that minimizes total travel time through a TSP-like heuristic using A* distances (Farrell et al., 17 Sep 2025).

The A* objective is described as time-optimal path planning formulated as shortest-distance trajectories on a grid-based roadmap. The planner uses the standard form

$f(n) = g(n) + h(n),$

with a distance-plus-orientation heuristic

$$h(s,g)=\|(x,y)-g\|_2 + \lambda \cdot \Delta\theta/\pi.$$

In hardware, $\lambda = 1$. In simulation, two variants are evaluated: $\mathbf{r}_c = K^{-1}\begin{bmatrix}u\v\1\end{bmatrix},$0 with $\mathbf{r}_c = K^{-1}\begin{bmatrix}u\v\1\end{bmatrix},$1 and $\mathbf{r}_c = K^{-1}\begin{bmatrix}u\v\1\end{bmatrix},$2 with $\mathbf{r}_c = K^{-1}\begin{bmatrix}u\v\1\end{bmatrix},$3. A higher $\mathbf{r}_c = K^{-1}\begin{bmatrix}u\v\1\end{bmatrix},$4 biases the planner toward safer, better-aligned trajectories at the possible cost of longer paths (Farrell et al., 17 Sep 2025).

Planning is continuously updated online. Replanning is triggered by new victim detections, map updates from the terrain-scouting UAV, UGV immobilization, or safety-driver intervention. The roadmap is persistent but updated, so A* is re-run from the UGV’s current position to the revised goals, and terrain-scouting UAV waypoints are updated to maintain the lead offset on the new forecast route. The planner therefore combines a mid-level global horizon, enabled by aerial cues, with local LiDAR-based safety enforcement, including emergency braking when nearby obstacles are too close (Farrell et al., 17 Sep 2025).

The framework’s distinction from a purely local baseline is explicit. The Local setting uses only a $\mathbf{r}_c = K^{-1}\begin{bmatrix}u\v\1\end{bmatrix},$5 LiDAR window, which produces myopic behavior in non-convex scenes. GLIDE instead uses a larger map with approximately 150 m range guided by UAVs. This long-horizon aerial guidance is the mechanism by which the UGV can anticipate structures such as U-shaped corridors rather than entering them reactively and becoming trapped (Farrell et al., 17 Sep 2025).

5. Hardware implementation and empirical performance

The hardware platform is unusually large for a laboratory SAR demonstration. The UGV, a GEM e6 golf cart, has maximum speed $\mathbf{r}_c = K^{-1}\begin{bmatrix}u\v\1\end{bmatrix},$6, steering rate $\mathbf{r}_c = K^{-1}\begin{bmatrix}u\v\1\end{bmatrix},$7, steering acceleration $\mathbf{r}_c = K^{-1}\begin{bmatrix}u\v\1\end{bmatrix},$8, PACMOD drive-by-wire actuation, GPS-RTK with NTRIP corrections, an Intel i9-14900K, 128 GB RAM, dual RTX 4080 GPUs, 4 TB storage, a Velodyne VLP32 LiDAR, and rate-limited PID trajectory tracking. Each Holybro X500 UAV runs PX4 firmware, has maximum speed $\mathbf{r}_c = K^{-1}\begin{bmatrix}u\v\1\end{bmatrix},$9, ascent and descent rate $\mathbf{r}_w = R(q)\mathbf{r}_c,$0, tilt limit $\mathbf{r}_w = R(q)\mathbf{r}_c,$1, thrust-to-weight ratio 2, a Jetson Orin Nano, a DC-DC converter, a nadir-facing OakD-Pro-W RGBD camera, an outer-loop PID for waypoint-to-velocity control, and inner-loop attitude stabilization by PX4 (Farrell et al., 17 Sep 2025).

Real-world evaluation used two outdoor scenarios: a U-shaped corridor formed by parked vehicles, with the UGV starting north and a victim south of the obstacle opening, and a linear east–west barrier of vehicles or obstacles with the same north–south start/goal arrangement. Three planning settings were compared: GT, in which the UGV has the full obstacle map; Local, in which the UGV relies only on a $\mathbf{r}_w = R(q)\mathbf{r}_c,$2 LiDAR window; and GLIDE, in which the terrain-scouting UAV flies 15 m ahead and provides traversability updates. The experiments used 10 runs per scenario per setting, randomized the start pose by $\mathbf{r}_w = R(q)\mathbf{r}_c,$3 m, and terminated trials either when the goal was reached or when a safety driver intervened (Farrell et al., 17 Sep 2025).

In hardware, the U-shaped scenario produced the following results: GT achieved 43.9 s, 102.40 m, and 100% success; Local achieved 51.27 s, 117.8 m, and 20% success; GLIDE achieved 47.17 s, 106.20 m, and 100% success. In the linear barrier scenario, GT achieved 44.70 s, 87.59 m, and 100% success; Local achieved 56.70 s, 146.45 m, and 60% success; GLIDE achieved 50.29 s, 108.73 m, and 100% success. The reported interpretation is that long-horizon UAV-guided planning is especially important in non-convex environments, where purely local planning falls into local minima or requires safety intervention (Farrell et al., 17 Sep 2025).

Simulation in PX4–Gazebo Harmonic with a Clearpath Warthog UGV and PX4 X500 UAV isolates the planning stack under perfect perception. Two world classes, easy and hard, were generated with U-shaped corridors and linear barriers, and each condition was run for 40 trials per scenario per setting. In easy worlds with $\mathbf{r}_w = R(q)\mathbf{r}_c,$4, GT obtained 50.08 s, 66.89 m, 100%; Local obtained 59.25 s, 92.10 m, 70%; GLIDE obtained 51.28 s, 74.55 m, 100%. In hard worlds with $\mathbf{r}_w = R(q)\mathbf{r}_c,$5, GT obtained 52.54 s, 73.76 m, 100%; Local obtained 63.50 s, 105.64 m, 65%; GLIDE obtained 59.37 s, 84.23 m, 90%. For $\mathbf{r}_w = R(q)\mathbf{r}_c,$6 in hard worlds, GT reached 57.60 s, 81.00 m, 95%; Local, 64.98 s, 108.91 m, 75%; and GLIDE, 59.82 s, 91.43 m, 95%. The larger orientation penalty therefore produced more conservative but safer paths, increasing GLIDE path length by approximately 7.2 m relative to $\mathbf{r}_w = R(q)\mathbf{r}_c,$7 while improving success rate by 5% in hard worlds (Farrell et al., 17 Sep 2025).

6. Limitations, extensions, and position within adjacent research

The reported limitations are primarily architectural rather than incidental. First, perception remains noisy in the real world because of GPS error, attitude variation, lighting, terrain texture, and wind. The system mitigates this with roll and pitch thresholds, image-center filtering, and consensus logic, but some georegistration error remains. Second, the terrain-scouting UAV currently validates only the UGV’s planned path with a fixed lead offset; it does not perform proactive or coverage-optimized exploration to discover shorter detours. The paper notes that in simpler obstacles, such as the line barrier, this can produce sparser coverage and more conservative paths. Third, the georeferencing model assumes a ground plane and the planner operates on a primarily 2D binary traversability map rather than soft-cost terrain models. Fourth, collision avoidance among UAVs and between UAVs and the UGV is relatively simple, relying largely on altitude separation and task structure. More complex multi-UAV and multi-UGV coordination is not yet implemented (Farrell et al., 17 Sep 2025).

These constraints delimit what GLIDE is and what it is not. It is not presently a proactive aerial exploration system, not a fully 3D terrain reasoner, and not a probabilistic planner over uncertain semantic maps. A plausible implication is that its strongest contribution lies in the middle layer between pure local navigation and fully autonomous global exploration: a role-separated, communication-efficient mechanism for injecting long-horizon structure into ground navigation under SAR timing constraints.

Several adjacent research strands clarify this position. “Drones Guiding Drones” studies a heterogeneous two-UAV framework in which a LiDAR-equipped primary UAV offloads dense mapping, global localization, and path planning to a less-equipped secondary UAV in cluttered GNSS-denied environments, using visibility-aware guiding viewpoints and periodic path transmission (Pritzl et al., 2023). “Cooperative Guidance for Aerial Defense in Multiagent Systems” addresses a different escort problem—defender protection of a high-value aerial evader in planar contested airspace—using time-constrained true proportional navigation and fixed-time convergence of engagement errors (Bajpai et al., 2 Oct 2025). “Autonomous 3D Moving Target Encirclement and Interception with Range measurement” develops a 3D guardian-drone architecture based on range-only localization, anti-synchronization encirclement, and zone-based switching between escort, monitor, and intercept modes (Liu et al., 16 Jun 2025). “Toward Increased Airspace Safety: Quadrotor Guidance for Targeting Aerial Objects” examines monocular line-of-sight guidance, true proportional navigation, and hybrid heading control for quadrotor interception of aerial targets (Bhattacharya, 2021). “FM-Planner” places foundation-model reasoning at the global path-planning layer for drones, combining an LLM with a vision encoder and validating low-frequency global planning with conventional high-frequency control in real flight (Xiao et al., 27 May 2025).

Taken together, these neighboring works suggest that GLIDE belongs to a broader class of guided autonomy systems in which information asymmetry is resolved through explicit role allocation. In GLIDE’s case, that asymmetry is aerial global sensing versus ground physical access. The framework’s future directions therefore follow directly from its present structure: adaptive lead offsets for terrain scouts, proactive detour discovery, online search-planning for the goal-searching UAV, richer traversability semantics, and more complex multi-agent coordination, all while preserving the core insight that long-horizon aerial guidance can materially improve time-critical ground SAR in unknown non-convex environments (Farrell et al., 17 Sep 2025).