Cheng-Lin (Stanley) Hsieh

Cheng-Lin (Stanley) Hsieh謝承霖

I’m a Ph.D. candidate in Computer Science and Engineering at the University of Michigan, advised by Stella X. Yu. I received my B.S. in Electrical Engineering from National Taiwan University. I study robot learning and visual perception in real-world settings.

Robot learning. I am interested in scalable learning from diverse data and experience, with human motion as one example of a useful prior. My current work focuses on contact-rich interaction and humanoid whole-body control.

Perception. I study spatial and dynamic understanding from partial observations, semantic localization, and fine-grained understanding of embodied video.

clhsieh@umich.edu

Portrait of Cheng-Lin Hsieh

Seeking summer 2027 research internships in robotics and computer vision.

Selected research

* Equal contribution

A frame of a girl and a dragon turns into DACE's predicted optical flow along a diagonal; a magnifier on the axe blade shows learned correspondence units following its edge.
Image to dense flow; the magnifier shows learned units along the blade’s edge.

NeurIPS 2026

Dense Flow from Adaptive Correspondence

Cheng-Lin Hsieh and Stella X. Yu

Optical flow must share evidence across pixels without mixing different motions. Instead of matching fixed grid cells, we learn the units of correspondence: each unit gathers evidence from the pixels it represents and returns its motion to them, with the largest error reductions where no match is visible and near motion boundaries.

Project page ↗arXiv Coming soonCode ↗Paper (PDF) ↗

Under review

Test-Time Motion Steering for Perceptive Humanoid Fall Recovery

Osher Azulay*, Cheng-Lin Hsieh*, and Stella X. Yu

We turn getting back up into a behavior the robot can adapt and steer. A visual diffusion model lets us guide heading, timing, and route at test time, while a depth-aware controller carries out the recovery on a real Unitree G1.

Project page ↗arXiv ↗Code Coming soon

Other publications

  1. Where to Hitch? Finding Small, Rare Parts Across Long Spatiotemporal Contexts

    C.-L. Hsieh*, Z. Wang*, S. X. Yu, E. Corona, S. Sasidharan, and E. D. Sumer.

    NeurIPS Workshop on Long-Context Foundation Models (LCFM), 2026.

    Finding a tiny part in a long video can be surprisingly difficult. HitchIt chooses useful views and tracks visual examples to locate parts such as trailer hitches without task-specific training.

    Paper (PDF) ↗Workshop page ↗

  2. EECS: Efficient End-Cloud Collaborative System with On-Demand Offloading Mechanism for Convolutional Neural Network ↗

    Y.-C. Lo, C.-L. Hsieh, H.-W. Chiang, and A.-Y. A. Wu.

    IEEE Internet of Things Journal, 12(8):11081–11095, 2025.

  3. Trainable Policy for the On-Demand Offloading on Edge-Cloud Collaborative System

    C.-L. Hsieh, Y.-C. Lo, and A.-Y. A. Wu.

    IEEE International Workshop on Machine Learning for Signal Processing (MLSP), 2023.

  4. Constraints-Aware Trainable Pruning with System Optimization for the On-Demand Offloading Edge-Cloud Collaborative System

    Y.-C. Lo, C.-L. Hsieh, and A.-Y. A. Wu.

    IEEE International Workshop on Machine Learning for Signal Processing (MLSP), 2023.

Teaching

Graduate Student Instructor · University of Michigan