Xianzheng Ma

World Models · Spatial Intelligence · Robot Learning

Xianzheng Ma

DPhil student, University of Oxford · Visual Geometry Group

From Modeling the World to General-Purpose Robot Intelligence

I study how robots can understand the physical world and generalize to new tasks.

Portrait of Xianzheng Ma
01

About

I’m a final-year DPhil student (since October 2023) in the Department of Engineering Science at the University of Oxford, supervised by Prof. Victor Prisacariu and Prof. Iro Laina, and part of both the Visual Geometry Group and the Active Vision Group. I also work closely with Prof. João F. Henriques and Prof. Ingmar Posner.

My research agenda connects perception, spatial understanding and physical world models, towards general-purpose robots that generalize:

  • Reliable perception. Understanding scenes under changing weather and visual domains.
  • Genuine spatial understanding. Testing whether 3D and language models truly reason about spatial relationships.
  • Physical world models & robotic interaction. Learning how the physical world evolves from video and 3D, to support robot understanding and interaction.
02

Latest News

Updated 2026-10-02

  1. Preprint

    Physis-Lang is now on arXiv · project page · code

  2. Award

    PhysVid: Best Paper Runner-Up, CVPR 2026 VGBE Workshop

  3. Award

    Real-3DQA: Best Paper Runner-Up, CVPR 2026 3D-LLM/VLA Workshop

  4. Grant

    Awarded the Tencent Qingyun Student Travel Grant for ICLR 2026.

  5. Accepted

    Our PhysVid paper has been accepted to CVPR 2026 Workshop VGBE.

Earlier news12 entries · 2023-10-02 → 2026-03-13
  1. Announced

    I will be attending ICLR 2026 in Rio de Janeiro, Brazil to present our Real-3DQA paper. Feel free to reach out if you’d like to connect!

  2. Accepted

    Our Beyond Single Object paper has been accepted to CVPR 2026 as a Finding.

  3. Accepted

    Our See It, Say It, Sorted paper has been accepted to CVPR 2026.

  4. Accepted

    Our Real-3DQA paper has been accepted to ICLR 2026.

  5. Release

    We have released the second version of our 3D-LLM survey paper, updated to include literature up to July 2025.

  6. Preprint

    New preprint: our PhysVid paper is now available on arXiv.

  7. Recognition

    Selected as an Outstanding Reviewer (top 3%) at ICCV 2025.

  8. Accepted

    Our Robotic Visual Instruction paper has been accepted to CVPR 2025.

  9. Preprint

    Check out our 3D-LLM survey paper on 3D-related tasks empowered by LLMs and other foundation models.

  10. Release

    We curated the Awesome-LLM-3D paper list for 3D-related tasks empowered by LLMs.

  11. Accepted

    Two papers are accepted in AAAI 2024.

  12. Started

    Started a new journey as a DPhil (PhD) student at the Visual Geometry Group and Active Vision Group, University of Oxford.

03

Milestones

2022 → 2026

  1. CVPR Oral as first author

    Foggy scene understanding; also REMOTE at AAAI and Rainy WCity at IJCAI.

  2. AAAI Outstanding Student Paper

    Decorate the Newcomers — plus a CVPR 2023 Highlight for Matching Is Not Enough.

  3. Started my DPhil at Oxford

    Visual Geometry Group & Active Vision Group.

  4. The 3D-LLM survey

    Mapping 3D tasks via multimodal LLMs; v2 in Oct 2025.

  5. Into robotics

    Robotic Visual Instruction accepted to CVPR 2025.

  6. ICCV Outstanding Reviewer

    Top 3% of reviewers at ICCV 2025.

  7. ICLR, CVPR & two runner-up awards

    Real-3DQA at ICLR 2026, with a Tencent Qingyun Student Travel Grant; two CVPR 2026 papers; Best Paper Runner-Up for Real-3DQA and PhysVid at CVPR 2026 workshops.

  8. Physis-Lang

    Physically plausible video world models, now on arXiv.

04

Research

2022 → Now

From robust perception to spatial understanding and physical world models — towards robots that generalize.

R.01 · 2022 – 2023

Reliable perception

Perception that holds up under fog, rain and domain shift; pose estimation with few labels.

Figure from Both Style and Fog Matter: adapting semantic segmentation from clear to foggy scenes

CVPR 2022OralFirst author

Both Style and Fog Matter

Adapting scene understanding to fog and changing visual styles.

paper ↗details
More in this group

AlsoREMOTE (AAAI 2022, first author) · Rainy WCity (IJCAI 2022, co-first author)

R.02 · 2023 – 2026

Genuine spatial understanding

Do 3D and multimodal models truly reason about space, or only appear to?

ICLR 2026Co-first author

Real-3DQA

Testing whether 3D language models truly understand spatial relationships.

Try live demo ↗ Project page ↗ Paper ↗

data ↗code ↗recorded talkdetails

More in this group

AlsoBeyond Single Object (CVPR 2026, Finding) · Point-Bind & Point-LLM (arXiv 2023)

R.03 · 2025 – now · current focus

Physical world models & robotic interaction

Can models learn how the physical world evolves — well enough to be useful to robots?

arXiv preprint · 2026Co-first author

Physis-Lang

Learning self-evolving physical language for more plausible video world models.

Project page ↗ Paper ↗ Code ↗

full gallery ↗details

More in this group
Full publication archive · 13 papers2022 – 2026 · complete authors, venues, awards, links

* equal contribution · † corresponding author · ‡ equal supervision · R.0n = research group

2026

  1. Physis-Lang: Self-Evolving Language as a Physical Representation for Video World ModelR.01

    Liming Lu*, Xianzheng Ma*, Wenkun He*, Guanqi Zhan*†, Yilin Zhao, Junyu Chen, Mengyao Xu, Jiaojiao Fan, Wenhang Ge, Yuchao Gu, Yunze Liu, Boyi Li, Zhen Dong, Victor Prisacariu, Ming-Yu Liu, Song Han, Han Cai†

    arXiv preprint, 2026

    paperproject pagecodebibtex

  2. Do 3D Large Language Models Really Understand 3D Spatial Relationships?R.02

    Xianzheng Ma*, Tao Sun*, Shuai Chen, Yash Bhalgat, Jindong Gu†, Angel X Chang, Iro Armeni, Iro Laina, Songyou Peng‡, Victor Adrian Prisacariu‡

    International Conference on Learning Representations (ICLR) 2026; CVPR 2026 3D-LLM/VLA Workshop Best Paper Runner-Up Award

    paperproject pagedatacode

  3. See It, Say It, Sorted: An Iterative Training-Free Framework for Visually-Grounded Multimodal Reasoning in LVLMsR.02

    Can each reasoning step be checked against visual evidence?

    Yongchang Zhang*, Xianzheng Ma*, Tianyi Liu, Guangquan Zhou, Yang Chen

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026

    paperpdfcodebibtex

  4. Beyond Single Object: Learning 3D Relations with Large Language ModelsR.02

    Kohsuke Ide, Ryousuke Yamada, Yue Qiu, Xianzheng Ma, Yoshihiro Fukuhara, Hirokatsu Kataoka, Yutaka Satoh

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026 (Finding)

    papercodebibtex

  5. Inferring Dynamic Physical Properties from Video Foundation ModelsR.01

    Can elasticity, viscosity and friction be read from video?

    Guanqi Zhan*, Xianzheng Ma*, Weidi Xie, Andrew Zisserman

    CVPR 2026 Workshop on Video Generative Models: Benchmarks and Evaluation (VGBE) Best Paper Runner-Up · arXiv preprint released 2025

    paperpdfcode

2025

  1. Robotic Visual InstructionR.01

    How should people tell a robot what to manipulate?

    Yanbang Li, ZiYang Gong, Haoyang Li, Xiaoqi Huang, Haolan Kang, Guangpingbai, Xianzheng Ma

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2025

    paperproject pagedataset

2024

  1. When LLMs step into the 3D World: A Survey and Meta-Analysis of 3D Tasks via Multi-modal Large Language ModelsR.02

    Where do LLMs stand on 3D understanding, generation and embodied tasks?

    Xianzheng Ma*, Brandon Smart*, Yash Bhalgat*, Shuai Chen, Xinghui Li, Jian Ding, Jindong Gu, Dave Zhenyu Chen, Songyou Peng, Jia-Wang Bian, Philip H Torr, Marc Pollefeys, Matthias Nießner, Ian D Reid, Angel X. Chang, Iro Laina, Victor Adrian Prisacariu

    IJCV under review · arXiv 2024, v2 2025

    paperarXivproject page / code

2023

  1. Point-Bind & Point-LLM: Aligning Point Cloud with Multi-modality for 3D Understanding, Generation, and Instruction FollowingR.02

    Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, Pheng-Ann Heng

    arXiv preprint, 2023

    papercode

  2. Matching Is Not Enough: A Two-Stage Framework for Category-Agnostic Pose EstimationR.03

    Can pose be estimated for object categories not seen in training?

    Min Shi, Zihao Huang, Xianzheng Ma, Xiaowei Hu, Zhiguo Cao

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2023 Highlight

    papercode

  3. Decorate the Newcomers: Visual Domain Prompt for Continual Test Time AdaptationR.03

    How can a model keep adapting as test domains keep changing?

    Yulu Gan, Xianzheng Ma, Yan Bai, Yihang Lou, Renrui Zhang, Nian Shi, Lin Luo

    AAAI Conference on Artificial Intelligence (AAAI) 2023 Outstanding Student Paper

    paper

2022

  1. Both Style and Fog Matter: Cumulative Domain Adaptation for Semantic Foggy Scene UnderstandingR.03

    Xianzheng Ma, Zhixiang Wang, Yacheng Zhan, Yinqiang Zheng, Zheng Wang, Dengxin Dai, Chia-Wen Lin

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2022 Oral

    paper

  2. REMOTE: Reinforced Motion Transformation Network for Semi-supervised 2D Pose Estimation in VideosR.03

    Xianzheng Ma, Hossein Rahmani, Zhipeng Fan, Bin Yang, Jun Chen, Jun Liu

    AAAI Conference on Artificial Intelligence (AAAI) 2022

    paper

  3. Rainy WCity: A Real Rainfall Dataset with Diverse Conditions for Semantic Driving Scene UnderstandingR.03

    Xian Zhong*, Xianzheng Ma*, Shidong Tu, Kui Jiang, Wenxin Huang, Zheng Wang

    International Joint Conference on Artificial Intelligence (IJCAI) 2022

    paper

Full, automatically updated list: Google Scholar ↗

05

Talks

Recording & slides

06

Academic service

Reviewing & service

  • Conference reviewerCVPR 2022–2026, AAAI 2022–2026, ICLR 2023–2026, NeurIPS 2024–2025, ICCV 2025, ECCV 2024
  • Journal reviewerIJCV, TPAMI, TMM
  • Area ChairIJCAI 2022 (China branch)

Workshops

07

Writing · 小红书

In Chinese

Xianzheng的具身学习笔记 @Oliver_embodied

Open profile ↗
  1. 01

    3D视觉还值不值得做？视频是更光明的未来吗？

    Read note ↗
  2. 02

    从谢赛宁访谈看恺明如何持续做出最顶的研究

    Read note ↗
  3. 03

    为什么VGG这么多Best Paper? 说说我的观察

    Read note ↗

Contact

For suggestions or collaborations

© Xianzheng Ma 马宪政 Content last updated 2026-10-15