README.md

🛸 LeDrone

State-of-the-art machine learning for real-world FPV drones : datasets, pretrained policies, simulators, and open hardware.

LeDrone brings the LeRobot formula to aerial robotics: a standardized dataset format, a policy zoo, native simulator integration, and affordable reference hardware. Lower the barrier to entry for learning-based drone flight, so everyone can contribute and benefit from shared datasets and pretrained models.

🤗 Built on LeRobot • Apache 2.0 • Simulation-first, safety-first (<250 g, EU class C0)

What you'll find here

• 📊 Datasets : LeDroneDataset format, extending LeRobotDataset v3.0: multi-rate aligned IMU (up to 1 kHz), FPV video (H.265, 60–120 fps), full FC telemetry, and human pilot RC inputs. Includes conversions of public benchmarks (UZH-FPV) and community-collected flights with automatic face blurring.
• 🧠 Pretrained policies : compact vision-based control models (CTBR action space, 60 Hz+), trained with a teacher-student + residual-model sim-to-real pipeline (Aerial Gym → real).
• 🕹️ Simulators : native integration with parallel RL simulators (Aerial Gym, Pegasus).
• 🔧 Open hardware : two reference builds: LeDrone-Lite (~300 €, ArduPilot + MAVLink, ground-station control, education & data collection) and LeDrone-Pro (500–600 €, Pi 5 / Orin Nano + Pixhawk, onboard inference & VIO).

Roadmap

1. Sim-only MVP : dataset format, reference datasets, ACT/SmolVLA baselines ✅ in progress
2. Hardware + collection : validated BOMs, MAVLink/MSP drivers, 50–100 h of annotated flight on the Hub
3. Sim-to-real : indoor hover transfer, then gate racing; pretrained models downloadable and flyable
4. Ecosystem : VLA extensions, workshop, native simulator support

Contribute

We're looking for ML researchers (policies, sim-to-real), embedded/FPV engineers (MAVLink/MSP, BOM validation), FPV pilots (data collection), and academic labs (scientific validation).

👉