OpenFlyV2: An onboard vision-language-action model for aerial perception and navigation
OpenFlyV2 reached 48% navigation success in four real-world environments without real-world aerial training data.
The model transferred directly from simulation to physical quadrotor flight. Running entirely onboard an NVIDIA Jetson Orin NX, it generated chunks of 12 flight waypoints at 10 Hz. A shared vision-language representation supports semantic understanding, language-grounded perception, spatial reasoning and continuous flight control. The simulation dataset contains approximately 70,000 navigation trajectories and 60,000 vision-language instruction–response pairs.
HF Daily Papers' note
The model transferred directly from simulation to physical quadrotor flight. Running entirely onboard an NVIDIA Jetson Orin NX, it generated chunks of 12 flight waypoints at 10 Hz. A shared vision-language representation supports semantic understanding, language-grounded perception, spatial reasoning and continuous flight control. The simulation dataset contains approximately 70,000 navigation trajectories and 60,000 vision-language instruction–response pairs.
HF Daily Papers' note
score 5