Yufei Ding | 丁昱菲

UniLatent reuses robot demonstrations across embodiments by aligning visual observations in a shared latent space rather than translating pixels. Alignment is supervised entirely by simulation renderings, requiring no target-robot demonstrations. In experiments, policies trained on one arm transfer zero-shot in simulation and few-shot on real hardware.

We introduce a comprehensive robotics benchmark, RBench, designed to evaluate robot-oriented video generation across five task domains and four distinct embodiments.

We introduce semantic orientation and propose a multi-modal model that facilitates spatial understanding in robot manipulation.

We present Open6DOR, a benchmark and zero-shot method for open-instruction 6-DoF object rearrangement, integrating 3D awareness to enable complex manipulation.

We propose RoboVerse, a comprehensive framework for advancing robotics through a simulation platform, synthetic dataset, and unified benchmarks.

We build an open platform that gathers comprehensive information of robotics and physics simulators for cutting-edge robot learning research.