A drone that takes plain-language directions and finds its own way there, without phoning home to a server, just flew 15 test missions indoors.
The stack, called VLN on the Fly, splits the job into separate stages instead of training one network to do everything. A quantized vision-language model matches an instruction to a rough image region, depth data turns that into a 3D goal point, a fast B-spline planner draws a flyable path, and a pretrained reinforcement-learning policy converts the path into motor commands. Researchers ran the stack on quadrotors across 15 flights toward three everyday objects in a controlled indoor space, reaching the target in 13 of 15 attempts with a mean goal error of 5.72 cm and average GPU use of 39.3%. Six more flights through cluttered obstacles stayed collision-free thanks to onboard perception gating.
The interesting part is the architecture choice, not the flight count. Most recent language-following robots fuse grounding, planning, and control into one end-to-end network, which is simpler to train but nearly impossible to debug when something goes wrong mid-flight. Keeping the stages separate means a failure in language grounding can be isolated from a failure in trajectory tracking, which matters more for a drone than for a robot that can just stop and idle on the ground.
Fifteen flights in a controlled room with three known objects is a proof of concept, not a field trial. Wind, GPS-denied navigation, and unfamiliar instructions are the tests that will actually decide if this design survives outside the lab.