WNM-3D: A World Navigation Model with 3D Scene Conditioning for Closed-Loop VLN
The episode discusses WNM-3D, a world navigation model that jointly predicts future video frames and actions for vision-language navigation. It uses 3D scene conditioning from monocular RGB history via a frozen geometry encoder, and a three-stage training pipeline with DAgger and DanceGRPO. Results show significant gains over baselines, with geometry conditioning and training order being key.