Mingdong Wu   |   吴铭东

I have been a Ph.D. student since 2021 in the School of Computer Science at Peking University, advised by Prof. Hao Dong. I received my bachelor degree in 2021, from Turing Class in Peking University.

My research interest lies in robot learning, generative models, reinforcement learning and 3D perception problems. I aim to build intelligent embodied agents that automatically discover and perform tasks without human specification.

Email  /  Google Scholar

We explore a pure generative approach to tackle the multi-hypothesis issue in 6D Category-level Object Pose Estimation. The key idea is to generate pose candidates using a score-based diffusion model and filter out outliers using an energy-based diffusion model. By aggregating the remaining candidates, we can obtain a robust and high-quality output pose.

The First Demand-driven Navigation Paper.

GFPose is a unified 3D human pose prior model that can be easily used for various applications, e.g., 3D human pose estimation, pose denoising and generation. Our key idea is to estimate the gradient field (a.k.a, score) of the perturbed human pose. We can leverage the gradient to adjust poses to be more plausible and feasible to a task specification.

We study object rearrangement without explicit goal specification. The agent is given examples from a target distribution and aims at rearranging objects to increase the likelihood of the distribution. Our key idea is to learn a target gradient field that indicates the fastest direction to increase the likelihood from examples via score-matching.