OpenFlyV2: An onboard vision-language-action model for aerial perception and navigation
OpenFlyV2: An onboard vision-language-action model for aerial perception and navigation
This paper introduces OpenFlyV2, a unified onboard vision-language-action model that enables semantic understanding, language-grounded perception, spatial reasoning, and continuous aerial control. The model generates flight waypoints at 10 Hz using an NVIDIA Jetson Orin NX and achieves 48% navigation success in real-world environments without prior aerial training data.
OpenFlyV2 demonstrates the feasibility of integrating vision, language, and action within a single model for aerial navigation, achieving zero-shot transfer from simulation to real-world flight with a 48% success rate across four environments.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Environment
- SimulationOpen source S4
Reported numbers
waypoints per action chunk
12
View original evidence
aypoints, while motion-aware temporal memory and action priors reduce the computational cost of closed-loop execution. We construct \dataset{}, a simulation-based dataset comprising approximately 70,000 navigation trajectories and 60,000 vision-language instruction--response pairs. OpenFlyV2 generates action chunks, each comprising 12 waypoints, entirely onboard at 10 Hz using an NVIDIA Jetson Orin NX. It transfers z
Open source S4action chunk generation rate
10
View original evidence
aypoints, while motion-aware temporal memory and action priors reduce the computational cost of closed-loop execution. We construct \dataset{}, a simulation-based dataset comprising approximately 70,000 navigation trajectories and 60,000 vision-language instruction--response pairs. OpenFlyV2 generates action chunks, each comprising 12 waypoints, entirely onboard at 10 Hz using an NVIDIA Jetson Orin NX. It transfers z
Open source S4
- code: Available Artifact link Open source S5
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
aypoints, while motion-aware temporal memory and action priors reduce the computational cost of closed-loop execution. We construct \dataset{}, a simulation-based dataset comprising approximately 70,000 navigation trajectories and 60,000 vision-language instruction--response pairs. OpenFlyV2 generates action chunks, each comprising 12 waypoints, entirely onboard at 10 Hz using an NVIDIA Jetson Orin NX. It transfers z
Open source S4
ero-shot from simulation to physical quadrotor flight, achieving a 48\% navigation success rate across four real-world environments without real-world aerial training data. Extensive results demonstrate that broad vision-language competence and continuous closed-loop flight can coexist within a single model on a resource-constrained aerial platform. Code will be made publicly available at https://github.com/Ideal-ljl
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org