A new robot navigation framework plans routes by predicting short videos of what the robot should see next.
The system, called CueNav, comes from a paper posted to arXiv. It pairs a video-generation model with two visual cues: a bird's-eye-view map that gives the robot global context, and a partial view of its own body kept visible in the first-person camera feed. A separate inverse-dynamics model then reads motion-flow fields out of the predicted video and converts them into actual robot commands, skipping the usual step of reconstructing a 3D scene from the footage. In maze tests, adding the map cue nearly doubled navigation success compared to video planning without it, and the body-aware setup hit 70 percent success squeezing through a narrow passage where comparison methods mostly failed.
Most video-model navigation work so far has leaned on short-horizon guidance and geometric reconstruction, which adds a lossy translation step between what the model imagines and what the robot actually does. CueNav's flow-based translation and embodiment-aware framing target that gap directly, and the claim that the same video planner worked across different robot platforms without retraining matters more than either success-rate number.
The benchmarks here are mazes and narrow passages, not real hallways full of people and clutter, so these numbers describe a lab ceiling, not a warehouse floor.