GLIDE: A Coordinated Aerial-Ground Framework for Search and Rescue in Unknown Environments

The paper presents GLIDE, a cooperative search-and-rescue framework that pairs two specialized UAVs with a UGV to enable rapid, safe navigation in unknown environments through real-time victim detection, terrain scouting, and guided long-horizon planning.

The Gist

Imagine a search-and-rescue mission in a disaster zone—like a collapsed building or a dense forest after a storm. The goal is simple: find the survivors and get them to safety as fast as possible. But the terrain is a nightmare: full of rubble, hidden pits, and dead ends.

This paper introduces GLIDE, a robotic team designed to solve this problem. Think of GLIDE not as a single robot, but as a highly coordinated rescue squad made up of three distinct characters, each with a specific job.

The Team: A "Ground & Air" Trio

1. The Ground Hero (The UGV):

   • Who: A sturdy, heavy-duty robot (in their experiment, a modified golf cart).
   • Job: It's the muscle. It can carry heavy medical supplies, push through debris, and physically reach the victim to help them.
   • The Problem: It's slow and "blind" to what's far away. If it tries to drive straight toward a victim, it might get stuck in a U-shaped dead end or fall into a hole it couldn't see. It only sees what's right in front of its bumper.

2. The "Spotter" Drone (Goal-Searching UAV):

   • Who: A fast, agile drone flying high above.
   • Job: It's the eyes in the sky. It scans the whole area like a hawk looking for a mouse. Its only job is to spot the survivors, figure out exactly where they are (using GPS coordinates), and shout down to the Ground Hero: "Hey! There's a person at these coordinates!"
   • Analogy: Think of this drone as a scout in a video game who marks the enemy on the mini-map so the player knows where to go.

3. The "Pathfinder" Drone (Terrain-Scouting UAV):

   • Who: Another drone, flying just ahead of the Ground Hero.
   • Job: It's the road inspector. While the Ground Hero is driving, this drone flies a few steps ahead to check the road. Is that pile of rocks actually passable? Is that puddle actually a deep hole? It sends a "traffic update" back to the Ground Hero.
   • Analogy: Imagine driving a car in thick fog. The Pathfinder drone is like a co-pilot who leans out the window, looks 50 feet ahead, and tells you, "Don't turn left, there's a wall!" before you even hit the brakes.

How They Work Together (The GLIDE Magic)

In the past, rescue robots often tried to do everything alone or relied on a human operator controlling them via a joystick. This is slow and risky, especially if the radio signal cuts out.

GLIDE changes the game by splitting the tasks:

1. The Hunt: The "Spotter" drone flies around, finds a victim, and sends a text message to the Ground Hero with the location.
2. The Plan: The Ground Hero gets the location and starts planning a route. But it doesn't just guess; it asks the "Pathfinder" drone to check the road ahead.
3. The Drive: As the Ground Hero drives, the Pathfinder drone is constantly updating the map. If the Ground Hero was about to drive into a dead end (a "local trap"), the Pathfinder drone sees it first and tells the Ground Hero, "Stop! Turn right instead."
4. The Result: The Ground Hero never gets stuck. It drives a smooth, safe, and fast path straight to the victim.

Why This Matters (The "Aha!" Moment)

The researchers tested this in two ways:

• Real Life: They used a golf cart and two drones in a park with fake obstacles (parked cars).
• Simulation: They ran thousands of computer tests.

The Results:

• Without the drones (The "Local" method): The Ground Hero acted like a person walking in the dark with a small flashlight. It often got stuck in dead ends or took huge, inefficient detours. It failed to reach the victim about 40-50% of the time in tricky situations.
• With GLIDE: The team reached the victim 100% of the time. They were almost as fast as if they had a perfect, magical map of the entire area (which humans don't have in real disasters).

The Big Takeaway

The paper proves that specialization works.

• If you ask one robot to do everything, it gets overwhelmed.
• If you let a human control everything, it's too slow and prone to error.
• But if you give the fast, flying robots the job of "looking ahead" and the slow, strong robot the job of "doing the heavy lifting," they become a super-team.

In simple terms: GLIDE is like giving a rescue team a GPS that updates itself in real-time and a co-pilot who never gets tired, ensuring that the heavy lifter never wastes a second driving into a dead end. This saves precious time, which in a disaster, literally saves lives.

Technical Summary

Here is a detailed technical summary of the paper "GLIDE: A Coordinated Aerial-Ground Framework for Search and Rescue in Unknown Environments."

1. Problem Statement

Search and Rescue (SAR) operations in unknown, unstructured environments (e.g., collapsed buildings, forests, flood zones) face significant challenges due to the inherent limitations of individual robotic platforms:

• Unmanned Aerial Vehicles (UAVs): Offer rapid mobility and wide-area coverage but suffer from limited flight endurance, payload capacity, and lack the ability to physically assist victims.
• Unmanned Ground Vehicles (UGVs): Provide persistence, heavy payload capacity, and physical interaction capabilities but are often constrained by "myopic" (short-sighted) local planning. They struggle to navigate non-convex environments or avoid large-scale obstacles without global situational awareness.
• Teleoperation Limitations: Current SAR deployments often rely on human teleoperation, which is bandwidth-intensive, prone to latency in signal-degraded environments, and cognitively overwhelming when managing multiple robots.

The core problem addressed is how to autonomously coordinate aerial and ground agents to achieve rapid victim localization and safe, time-efficient navigation in unknown environments without relying on continuous human intervention.

2. Methodology: The GLIDE Framework

The authors propose GLIDE (Guided Long-horizon Integrated Drone Escort), a cooperative multi-agent framework that explicitly decouples aerial roles to complement a ground ego-vehicle.

A. System Architecture

The system consists of three primary agents with distinct responsibilities:

1. Goal-Searching UAV:
   • Role: Conducts rapid site surveys to detect and geolocate victims.
   • Perception: Executes lightweight, real-time object detection (YOLO-based) onboard. It transmits only compact, georeferenced detection messages (bounding boxes + coordinates) to the ground vehicle to save bandwidth.
2. Terrain-Scouting UAV:
   • Role: Flies ahead of the UGV's planned trajectory (with a lead offset) to provide mid-level situational awareness.
   • Perception: Captures imagery to infer traversability (identifying non-traversable regions like rubble, water, or voids). It generates a "mid-level traversability map" to update the UGV's planning graph.
3. Ground Ego-Vehicle (UGV):
   • Role: The primary rescue agent responsible for navigating to victims.
   • Planning: Fuses victim coordinates from the Goal-Searching UAV and the traversability map from the Terrain-Scouting UAV with local LiDAR data. It performs long-horizon A path planning* and continuous replanning as new information arrives.

B. Technical Implementation

• Perception Stack:
  • Utilizes YOLOv11 fine-tuned on the VisDrone dataset (top-down aerial imagery) for victim detection.
  • Models are optimized with TensorRT and CUDA to run at 30 Hz on edge devices (Jetson Orin Nano).
  • Georeferencing is achieved by projecting 2D bounding boxes to the ground plane using camera intrinsics, GPS, and altitude data, with strict filtering for UAV orientation (roll/pitch < 15°) to ensure accuracy.
• Planning Stack:
  • The UGV uses an A algorithm* on a discretized grid (300x300, 0.5m resolution).
  • Heuristic: Uses a distance-plus-orientation heuristic ($h(s, g) = \|(x, y) - g\|_2 + \lambda \cdot \Delta\theta/\pi$) to favor paths that are both spatially close and properly aligned.
  • Replanning: The graph is persistent; non-traversable regions are treated as hard obstacles, and the plan is updated dynamically as the Terrain-Scouting UAV reveals new terrain data.
• Communication:
  • Uses a Wi-Fi backhaul (TP-Link Archer AX21).
  • Transmits only compact metadata (<1 kB/event) rather than raw video streams to handle bandwidth constraints.
  • Implements a queuing mechanism: if connectivity is lost, detections are cached onboard and forwarded upon reconnection.

3. Key Contributions

1. Explicit Role Separation: A novel framework that separates aerial tasks into "Goal-Searching" (victim detection) and "Terrain-Scouting" (obstacle mapping), allowing the ground vehicle to focus on execution and long-horizon planning.
2. Lightweight Edge Perception: A real-time detection pipeline optimized for resource-constrained UAVs, enabling onboard inference without streaming raw video.
3. Aerial-Guided Planning: An A* planning stack that integrates mid-level traversability maps from aerial scouts to generate time-efficient, collision-free trajectories for the UGV.
4. Hardware Validation: A proof-of-concept demonstration using a GEM e6 golf cart (UGV) and two Holybro X500 UAVs, validating the end-to-end system in real-world outdoor scenarios.
5. Simulation Ablations: Comprehensive simulation studies (PX4-Gazebo) that isolate the planning stack from perception to evaluate performance under varying environmental complexities.

4. Experimental Results

Hardware Experiments (Real-World)

The system was tested in two scenarios: a U-shaped corridor (non-convex obstacle) and a Linear barrier.

• Success Rate: The GLIDE framework achieved a 100% success rate in both scenarios. In contrast, the "Local" strategy (UGV relying only on onboard LiDAR) failed significantly (20% success in U-shaped, 60% in Linear) due to getting trapped in local minima.
• Efficiency:
  • U-Shaped: GLIDE took 47.17s (vs. 43.9s for Ground Truth), a negligible delay of ~3.3s.
  • Linear: GLIDE took 50.29s (vs. 44.7s for Ground Truth). The slight delay was attributed to the terrain scout spending less time hovering over simple linear obstacles, leading to sparser map updates.
• Conclusion: Explicit aerial guidance drastically improves navigation safety and success rates in unknown, non-convex environments compared to reactive local planning.

Simulation Experiments

Simulations were run with perfect detection assumptions to isolate the planning algorithm's performance.

• Performance Gap: GLIDE consistently narrowed the gap to the Ground Truth (GT) planner, significantly outperforming the Local strategy.
• Heuristic Tuning: Increasing the orientation penalty coefficient ($\lambda$) from 1.0 to 2.5 resulted in safer, more conservative paths. In "Hard" worlds, this increased success rates from 90% to 95% at the cost of slightly longer path lengths.
• Robustness: The simulation confirmed that the framework is robust to environmental complexity, provided the terrain scouts can update the map with sufficient frequency.

5. Significance and Future Work

Significance:
The paper demonstrates that autonomous cooperation between heterogeneous robots (UAVs and UGVs) is a viable solution for time-critical SAR missions. By offloading perception and situational awareness to aerial agents, the ground robot can navigate complex, unknown environments safely and efficiently without human teleoperation. This approach addresses the critical trade-off between the speed of aerial systems and the endurance of ground systems.

Future Work:

• Adaptive Search Planning: Implementing online algorithms to dynamically cover high-probability victim locations rather than relying on external waypoint inputs.
• Dynamic Lead Offsets: Adjusting the distance between the Terrain-Scouting UAV and the UGV dynamically based on UGV speed and environmental complexity to optimize the exploration-exploitation trade-off.
• Perception-Aware Planning: Developing planning strategies that explicitly account for sensor noise, latency, and incomplete map data to bridge the gap between simulation and real-world performance.