OpenFlyV2 Runs Full Vision-Language-Action Flight Model on Jetson Orin NX

A quadrotor running OpenFlyV2 flew without a single frame of real-world aerial footage in its training set, and landed a 48% navigation success rate across four physical environments. The model loads entirely onto an NVIDIA Jetson Orin NX, generating action chunks of 12 waypoints every tenth of a second.

The paper on Hugging Face describes OpenFlyV2 as "a unified onboard vision-language-action architecture that uses a shared vision-language representation to support semantic understanding, language-grounded perception, spatial reasoning, and continuous aerial control within a single model." A "compact action expert" turns that shared multimodal context into short-horizon flight waypoints, with "motion-aware temporal memory and action priors" cutting the compute cost of closed-loop execution.

Everything the model knows about flying it learned in a simulator. The authors assembled roughly 70,000 navigation trajectories and 60,000 vision-language instruction-response pairs, and the system then "transfers zero-shot from simulation to physical quadrotor flight."

The abstract reports the 48% figure as a single number across four environments; it publishes no per-environment breakdown and no failure analysis. OpenFlyV2 lands in a run of edge-deployed model work on our radar this week, next to a solo-dev TinyWorldformer squeezing a transformer world model onto a Raspberry Pi Pico 2 W.