ICARUS
A defense drone that hears a vehicle before it sees it, then identifies it, maps its weak points and plans the approach.
- Result
- 2nd place, Drone Defense Hackathon — Grand Palais, Paris, Nov 2025 · 28 teams · French Ministry of the Armed Forces · follow-up by the French Army for real-condition trials
- Hardware
- DJI Matrice 4E · NVIDIA Jetson · Arduino wake-box
- Stack
- TensorRT, ROS, MAVLink, RF-DETR, TensorFlow, Librosa, ByteTrack, A*/RRT-Connect
How it works
STANDBY mic ─► acoustic CNN ─► vehicle? ─► power on & take off
ACTIVE camera ─► RF-DETR + 3D pose ─► vulnerability map
─► human validation ─► A*/RRT waypoints ─► autopilot
- Acoustic wake-up. The drone stays off; a stand-alone box classifies Mel spectrograms with a 2D CNN (vehicle vs. environment) and powers the drone on. 92% / 96% accuracy; standby goes from <1 h to ~8 h.
- Vision. RF-DETR outputs class and 2D box; a regression head plus the vehicle's known dimensions recover a 3D box and pose.
- Vulnerabilities. Critical points of a canonical 3D model (turret, hatches, turret–hull junction) are projected into the scene every frame and scored.
- Approach. Sampling-based planner (A*/RRT-Connect) toward a stand-off disc around the target; waypoints to the flight controller. Human validation is required before any offensive action.
Results
| Acoustic wake-up | Predicted vehicle | Predicted environment |
|---|---|---|
| Actual vehicle | 92% | 8% |
| Actual environment | 4% | 96% |
- Standby autonomy: <1 h (drone on) → ~8 h (wake-box)
- Real-time on the Jetson with TensorRT: vision 10–15 FPS, planning 2–5 Hz
- Validated offline, in simulation, and end to end on a Leopard 2 scale model
