Xinrui Zu (俎新瑞)

Amsterdam, the Netherlands

Computational Intelligence Group

Quantitative Data Analytics Group

Xinrui Zu (祖新瑞) is a PhD student at Vrije Universiteit Amsterdam, jointly in the Computational Intelligence Group and the Quantitative Data Analytics Group, advised by Dr. Kevin Sebastian Luck and Dr. Shujian Yu. He studies the geometry of learned representations — how perception is structured inside neural networks — and uses optimal transport to align that structure across sensory modalities (vision, force, tactile, IMU, proprioception) and across domains (medical imaging and robotics).

The same idea threads through his work: a Gromov–Wasserstein regularizer for robot perception (UniOMA, ICRA 2026), optimal-transport flows for medical image synthesis, and embedding methods that preserve structure under dimensionality reduction (SpaceMAP, ICML 2022; Deep Recursive Embedding, IEEE TVCG 2022). Before his PhD he was a research assistant at TU Delft, Imaging Physics with Dr. Qian Tao, earned an MSc in Robotics & Mechatronics at the University of Twente, and a BSc in Theoretical & Applied Mechanics at Fudan University. And because alignment shouldn’t stay theoretical, he also builds the robots — modular manipulators, autonomous vehicles, and the sensors they learn from.

Feel free to reach out for collaborations or a chat about optimal transport, multimodal robot learning, or generative models.

Featured — Σ-LeWM (2026)

Σ-LeWM: forward, backward, and inverse dynamics jointly structure the latent world model for robot control.

Σ-LeWM: Learning World Models via S-Information for Robot Control combines forward, backward, and inverse dynamics to learn structured latent representations, improving robot control and planning efficiency. Explore the project →

What does Xinrui Zu work on?

Xinrui Zu develops optimal-transport methods — primarily Gromov–Wasserstein — that align the structure of different robot sensor modalities rather than matching them point-to-point, improving robustness when sensors fail or drop out. His broader interests span generative models for medical imaging and dimensionality reduction for high-dimensional data visualization.

news

Nov 09, 2026	New preprint: Σ-LeWM: Learning World Models via S-Information for Robot Control. We combine forward, backward, and inverse dynamics to learn latent world models for more effective and efficient robot planning. [Project] · [Code]
May 15, 2026	🎉 UniOMA was presented at ICRA 2026 Late-Breaking Results and accepted to the ICRA 2026 “From Data to Decisions” workshop (Poster Session 2), Vienna. [Paper]
Apr 10, 2024	cHeartFlow — synthesizing cardiac MR images from sketches — was accepted to MIDL 2024. [Paper]
May 15, 2022	🎉 SpaceMAP was accepted to ICML 2022. [Paper] · [Code] · [Video]
Jan 20, 2022	Deep Recursive Embedding was published in IEEE TVCG. [Paper] · [Code]

selected publications

Preprint
Σ-LeWM: Learning World Models via S-Information for Robot Control

Xinrui Zu, Michael Przystupa, Zhao Yang, and 2 more authors

Preprint, Nov 2026

Abs Bib HTML Code

Σ-LeWM augments LeWorldModel with backward dynamics and inverse dynamics objectives to structure the learned latent representation through S-information. We demonstrate improved control performance and planning efficiency on standard world-model benchmarks and simulated robotic manipulation tasks, while retaining the same forward planner at test time.
@article{zu2026sigmalewm, title = {Σ-LeWM: Learning World Models via S-Information for Robot Control}, author = {Zu, Xinrui and Przystupa, Michael and Yang, Zhao and Yu, Shujian and Luck, Kevin Sebastian}, journal = {Preprint}, year = {2026}, month = nov, }
ICRA’26
UniOMA: Unified Optimal-Transport Multi-Modal Structural Alignment for Robot Perception

Xinrui Zu, Kevin Sebastian Luck, and Shujian Yu

In IEEE ICRA 2026 Workshop “From Data to Decisions” (also Late-Breaking Results), 2026

Abs Bib HTML PDF

UniOMA uses Gromov–Wasserstein optimal transport to align the structure of heterogeneous robot sensors (vision, force, tactile, IMU, proprioception) into a shared representation space, scaling linearly to three or more modalities and improving robustness to sensor dropout. Contrastive objectives such as InfoNCE align modalities at the instance level but cannot preserve intra-modal geometry — a structural alignment gap that UniOMA closes with a learned Gromov–Wasserstein barycenter regularizer.
@inproceedings{zu2026unioma, title = {UniOMA: Unified Optimal-Transport Multi-Modal Structural Alignment for Robot Perception}, author = {Zu, Xinrui and Luck, Kevin Sebastian and Yu, Shujian}, booktitle = {IEEE ICRA 2026 Workshop ``From Data to Decisions'' (also Late-Breaking Results)}, year = {2026}, }
MIDL’24
cHeartFlow: Synthesizing Cardiac MR Images from Sketches

Xinrui Zu and Qian Tao

In Medical Imaging with Deep Learning (MIDL), 2024

Abs Bib HTML

A contrastive generative framework that synthesizes cardiac magnetic resonance (CMR) images from simple sketches by training on contrastive pairs of images and sketches, enabling controllable medical image synthesis and zero-shot registration.
@inproceedings{zu2024cheartflow, title = {cHeartFlow: Synthesizing Cardiac MR Images from Sketches}, author = {Zu, Xinrui and Tao, Qian}, booktitle = {Medical Imaging with Deep Learning (MIDL)}, year = {2024}, }
ICML’22
SpaceMAP: Visualizing High-Dimensional Data by Space Expansion

Xinrui Zu and Qian Tao

In International Conference on Machine Learning (ICML), 2022

Abs Bib HTML Code

A dimensionality-reduction method that visualizes data of any dimensionality on a 2-D map. We analytically derive a transformation of distance between high- and low-dimensional spaces to match their capacity, and show it provably reduces the intrinsic dimension of high-dimensional data within the maximum-likelihood intrinsic-dimensionality framework.
@inproceedings{zu2022spacemap, title = {SpaceMAP: Visualizing High-Dimensional Data by Space Expansion}, author = {Zu, Xinrui and Tao, Qian}, booktitle = {International Conference on Machine Learning (ICML)}, year = {2022}, }
TVCG’22
Deep Recursive Embedding for High-Dimensional Data

Zixia Zhou, Xinrui Zu, Yuanyuan Wang, and 2 more authors

IEEE Transactions on Visualization and Computer Graphics (TVCG), 2022

Abs Bib HTML Code

Deep Recursive Embedding (DRE) is a dimensionality-reduction method that leverages latent data representations for boosted embedding performance, maps out-of-sample data, and scales to extremely large datasets, improving both local and global structure preservation over the state of the art.
@article{zhou2022dre, title = {Deep Recursive Embedding for High-Dimensional Data}, author = {Zhou, Zixia and Zu, Xinrui and Wang, Yuanyuan and Lelieveldt, Boudewijn P. F. and Tao, Qian}, journal = {IEEE Transactions on Visualization and Computer Graphics (TVCG)}, year = {2022}, }