Combining diffusion model and generalizable optimizer to enhance urban scene 3DGS, enabling high-fidelity rendering at extreme views.

Combining multiple sensors and odometry output with state-space model to predict heteroscedastic localization-error distributions across time.

A method to detect object changes of the same scene in different time, enabling downstream tasks like robot tidying.

An unsupervised generalizable 4D Gaussian reconstruction and future forecasting framework for driving scenes.

A VLM-based robot navigation system that uses human cues to efficiently navigate in unknown human-built environments.

A framework that generates 4D-consistent and realistic LiDAR point clouds from an input HD map.