The Problem
Drones are useless without GPS. GPS jamming in Ukraine took out 85% of some squadrons. Indoors, urban canyons, forests — GPS simply fails.
Military solutions cost $50K+. Academic code is unmaintained. There's no real open-source answer for GPS-denied drone flight.
What I Built
GhostPilot — an open-source visual-inertial SLAM + agentic AI navigation stack that lets any drone fly without GPS.
Key Features
- Visual-Inertial SLAM: Camera + IMU fusion for 6DOF pose estimation — no GPS needed
-
Agentic Mission Planner: Natural language commands → executable navigation goals
- "Fly to the third floor, check each room for occupants, land at the helipad"
- "Inspect the roof, avoid personnel, resume path at waypoint B"
- Nav2 Integration: Industry-standard path planning + obstacle avoidance
- Edge-Native: Runs on Jetson Orin / Raspberry Pi 5, no cloud dependency
Quick Start
git clone https://github.com/amsach/GhostPilot.git
cd GhostPilot
./scripts/setup_jetson.sh
ros2 launch ghostpilot_gazebo indoor_warehouse.launch.py
ros2 run ghostpilot_agent mission_parser_node
Architecture
| Package | Description |
|---|---|
ghostpilot_core |
VINS-Mono SLAM + Nav2 integration |
ghostpilot_agent |
LLM-based mission parser + executor |
ghostpilot_gazebo |
Gazebo simulation world + models |
Hardware
- Compute: NVIDIA Jetson Orin AGX or Raspberry Pi 5
- Camera: Intel RealSense D435i (stereo + IMU)
- Frame: Any MAVLink-capable quadcopter
Comparison
| Feature | GhostPilot | Skydio | Military |
|---|---|---|---|
| Cost | $0 (open-source) | $5K+ | $50K+ |
| GPS-denied | ✅ Native | ⚠️ Limited | ✅ Yes |
| Agentic AI | ✅ Natural language | ❌ Waypoints | ❌ Pre-programmed |
| ROS2-native | ✅ Full | ❌ Closed | ❌ Proprietary |
Status
Core SLAM + Nav2 bridge working in simulation. Agentic layer in progress. PRs welcome.