OpenFlyV2 demonstrates onboard aerial VLA inference
The OpenFlyV2 authors describe a shared vision-language architecture with a compact action expert for flight waypoints and temporal memory. They report approximately 70,000 simulated navigation trajectories and 60,000 instruction-response pairs for training. On NVIDIA Jetson Orin NX, the system reportedly generates 12-waypoint chunks entirely onboard at 10 Hz. The authors claim zero-shot transfer to four real environments with 48% navigation success and no real-world aerial training data. This is a research result, not a certified product or cross-chip benchmark. Model size, power, memory use and latency distribution are undisclosed; code release is promised, not confirmed.