FalconGym: Photorealistic Simulation for Zero-Shot Sim-to-Real Vision-Based Quadrotor Navigation
Train in NeRF, fly in the real world: 95.8% success over 120 real gates, zero fine-tuning.
Train in NeRF, fly in the real world: 95.8% success over 120 real gates, zero fine-tuning.
Edit the scene in milliseconds, train where the policy is weakest: 98.6% on hardware, including a moving gate.
10,000 labeled images in 20 minutes, zero manual annotation; onboard physics-aware tracking at 25 Hz.
A 150-gram fixed-wing that lands itself, vision-only, trained entirely in a Gaussian-splat twin.
A real quadrotor through virtual obstacles at 28 ms latency: crash-free failure-boundary testing.

Two minutes of phone video to a trained perception model, then live inference on the phone.

Fly the photorealistic twin by hand, with every gate crossing scored in 3D.

Camera and IMU only, all seven gates. Distinguished Drone Operations Award.









