A new drone navigation model skips the expensive step of imagining full video frames before deciding where to fly.
Researchers built DiffWAM, a navigation system that pulls motion information straight out of a frozen video prediction model instead of generating complete future video and reconstructing 3D geometry from it. Two components, called Grid-Motion and Latent2Pose, preserve the model's sense of motion over time and anchor it to the geometry of the drone's first-frame view, turning that into a usable 3D flight path. On a 1,000-sample benchmark the team built for testing, DiffWAM hit a trajectory error of about 0.35 meters and landed within its target 74.4% of the time. A lighter version, DiffWAM-Flash, ran onboard an NVIDIA Jetson AGX Thor chip with 1.08 seconds of latency, and real-world flights included obstacle-weaving, orbiting, and multi-stage routes.
Most video-based world models for robots generate an imagined future video first, then reconstruct 3D structure from it before the robot can act - a slow, compute-heavy pipeline for anything moving in real time. DiffWAM's shortcut, converting prediction directly into motion, matters because it suggests large pretrained video models could steer real hardware without a rack of GPUs riding along.
The 74.4% endpoint success rate still means roughly one flight in four misses its mark, and the real-world demos are described as 'representative' rather than exhaustive - promising, but not yet proof this holds up outside curated tests.