Liu Liu

By structurally disentangling temporal joint planning from frame-level manipulation, StructBiHOI achieves superior long-horizon stability and motion realism in bimanual hand-object interaction generation.

Achieve robust MoE inference on analog hardware without retraining by intelligently offloading only the most noise-sensitive experts to digital compute.

Achieve state-of-the-art image fusion in just one minute of training, bridging the gap between slow deep learning methods and fast but less adaptable traditional techniques.

Skip the 3D modalities: Spa3R shows that strong spatial reasoning can emerge directly from 2D vision alone by learning to predict feature fields from multi-view images.

Ditch continuous pose regression for articulated objects: DICArt's discrete diffusion approach unlocks more reliable 6D pose estimation by respecting kinematic constraints and navigating the search space more effectively.