FATHOM: World-Model-Guided Diffusion Trajectory Selection for AUV Tracking

Xiaorui Wang and other authors propose FATHOM, a world-model-guided diffusion action trajectory selection framework for relative tracking of autonomous underwater vehicles (AUVs) under sparse active light markers. The method first uses a multimodal task-supervised recurrent state-space model to fuse marker observations, inertial measurements, action history, and temporal dynamics to form a predictive tracking belief, and then a diffusion action prior generates multiple multi-thruster trajectories. Rather than directly executing the generated trajectories, FATHOM selects the optimal candidate under an explicit receding-horizon objective and continuously updates decisions based on onboard observations. The authors report that in real AUV experiments, including comparisons with strong pose-based baselines and component ablations, FATHOM achieves robust autonomous tracking under normal and challenging motions and when markers naturally become partially visible, and maintains onboard real-time operation without relying on motion-capture feedback.

AI 根据报道生成 · 2 days ago更新