Researchers led by Zhexuan Zhou have built a way for drones to follow navigation instructions given in everyday language, according to a paper on arXiv. Such flights need the drone to follow the instruction closely, keep track of how much of it is done and avoid obstacles, which current world-action models handle poorly, the authors say.
Their method adds three parts: training that makes the drone's motion respond to the instruction, a prediction of how much of the instruction has been carried out so the drone knows when to stop, and training in scenes made of 3D Gaussians to improve obstacle avoidance. Producing each action in a single denoising step cuts the computing needed.
In closed-loop tests the drone completed more steps of its tasks, reached goals more safely and stopped at the right time. Real flights indoors and outdoors confirmed the approach. The abstract gives no figures.



