Yuxuan Xie

Real-time open-vocabulary segmentation for robotics is now possible, unlocking faster sim-to-real transfer with a 2x improvement in downstream navigation tasks.

Achieve real-time, high-quality panoptic 3D reconstruction by directly lifting 2D features into 3D space, bypassing traditional multi-stage pipelines.

Unlock individual-level causal reasoning in the real world, even with incomplete causal knowledge, using tighter bounds on probabilities of causation.