Pushi Zhang (张蒲石)

About Me

I am a researcher and AI engineer building intelligent robotic systems that perceive, understand, and interact with the physical world. Specifically, I work on embodied and physical AI: training foundation models that combine general multimodal understanding with reliable real-world action.

Currently, I work at X Square Robot, where I lead a critical part of our foundation model research and have been a core contributor to:

Wall-OSS-0.5 — our open-source 4B vision-language-action foundation model, whose pretrained checkpoint already completes real-robot manipulation tasks before any task-specific fine-tuning;
WALL-B logistics sorting — fully autonomous sorting of real-world parcels, deployed in live production;
TwinDEX — an embodiment-consistent, high-efficiency data collection system for dexterous hands.

I have also contributed to X-Tokenizer and WALL-WM.

Previously, I was a researcher at Microsoft Research Asia in Li Zhao’s group, where I worked on reinforcement learning and embodied foundation models (latent action models, VLA, visual navigation).

More broadly, I hope my research benefits the vast majority of people, making what physical intelligence can offer accessible to everyone.

Join Us

I hope to work with exceptional people who share the vision of physical AGI, as interns or full-time colleagues. I especially value:

Frontier AI fluency — deep command of frontier LLMs and agents, and the engineering to build agent-native pipelines that automate physical-AI research and data workflows;
Embodied data expertise — mastery of the full embodied data stack, from collection and quality control to training, built into rigorous, reproducible systems;
Research depth — genuine depth in a chosen area, and the judgment to identify and rigorously test the questions that matter on the path to physical AGI;
Strong foundations — solid mathematics and excellent engineering skills.

If you are interested, or would like to exchange ideas, feel free to reach out.

Get in touch Learn more →

Featured Work

Wall-OSS-0.5

An open-source 4B vision-language-action foundation model, pretrained across 20+ robot embodiments, that already completes real-robot manipulation tasks before any task-specific fine-tuning.

Paper Code Model Project

WALL-B Logistics Sorting

Our WALL-B embodied foundation model drives two robot arms to sort highly variable real-world parcels fully autonomously: 1,816 parcels per hour at over 98% accuracy in a public livestream, and now deployed on a live production line with a leading logistics company.

News Xinhua

TwinDEX

An embodiment-consistent, high-efficiency data collection system for dexterous hands: a wearable device and a matching robot hand share one design, so human demonstrations transfer directly to the robot at scale.

Project

villa-X

A visual-language-latent-action model that plans in imagination: from an image and an instruction, it predicts latent actions and the motion they lead to, bridging high-level planning and low-level robot execution.

Paper Code Model Project

X-Tokenizer

A multimodal action tokenizer that serves as a semantic interface between vision-language reasoning and continuous robot control.

Paper Code Model Project

WALL-WM

A world action model that learns from semantic action events rather than fixed-length action chunks.

Paper Code Project

News

Sep 2026 Introduced TwinDEX, an embodiment-consistent, high-efficiency data collection system for dexterous hands.
Aug 2026 Sorted real-world parcels fully autonomously with our WALL-B model in a global livestream: 1,816 parcels per hour at over 98% accuracy.
Jun 2026 Released X-Tokenizer, a multimodal action tokenizer for VLA pretraining, and WALL-WM, a world action model built on semantic action events.
May 2026 Open-sourced Wall-OSS-0.5, our 4B vision-language-action foundation model, with the technical report, code, and model weights.
Jan 2026 Our paper villa-X is accepted by ICLR 2026.
Sep 2025 Released code and models of PIG-Nav, which achieves strong performance in image-goal navigation.
Oct 2024 Released IGOR: Image-GOal Representations are the Atomic Control Units for Foundation Models in Embodied AI.

Selected Papers

* Equal contribution · Full publication list →

Report

Wall-OSS-0.5 Technical Report

Ryan Yu, Pushi Zhang, Starrick Liu, et al. 等 Ryan Yu, Pushi Zhang, Starrick Liu, Brae Liu, Miracle Kang, Shalfun Li, Lights Shi, Ellie Ma, Ping Yang, Chris Pan, Jerry Chen, Dongxiu Liu, Rain Sun, Miles Guo, Byron Zhang, Hugo Zhou, Zach Xu, Vincent Chen, Harrison Huang, James Wang, Dance Kuzi, Andy Zhai, Hang Su, Roy Gan, Lucy Liang, Hao Wang, and Qian Wang

Technical report, arXiv:2605.30877, 2026

arXiv Code Model Project

@article{yu2026walloss,
  title = {Wall-OSS-0.5 Technical Report},
  author = {Yu, Ryan and Zhang, Pushi and Liu, Starrick and Liu, Brae and Kang, Miracle and Li, Shalfun and Shi, Lights and Ma, Ellie and Yang, Ping and Pan, Chris and Chen, Jerry and Liu, Dongxiu and Sun, Rain and Guo, Miles and Zhang, Byron and Zhou, Hugo and Xu, Zach and Chen, Vincent and Huang, Harrison and Wang, James and Kuzi, Dance and Zhai, Andy and Su, Hang and Gan, Roy and Liang, Lucy and Wang, Hao and Wang, Qian},
  journal = {arXiv preprint arXiv:2605.30877},
  year = {2026},
}

Preprint

X-Tokenizer: A Multimodal Action Tokenizer for Vision-Language-Action Pretraining

Miracle Kang^*, Lights Shi^*, Lucy Liang, …, Pushi Zhang, et al. 等 Miracle Kang^*, Lights Shi^*, Lucy Liang, Roy Gan, Dongxiu Liu, Pushi Zhang, Sylas Chen, Shawn Qin, Yinan Zheng, Jinliang Zheng, Hao Wang, Xianyuan Zhan, and Hang Su

Preprint, arXiv:2606.14752, 2026

arXiv Code Model Project

@article{kang2026xtokenizer,
  title = {X-Tokenizer: A Multimodal Action Tokenizer for Vision-Language-Action Pretraining},
  author = {Kang, Miracle and Shi, Lights and Liang, Lucy and Gan, Roy and Liu, Dongxiu and Zhang, Pushi and Chen, Sylas and Qin, Shawn and Zheng, Yinan and Zheng, Jinliang and Wang, Hao and Zhan, Xianyuan and Su, Hang},
  journal = {arXiv preprint arXiv:2606.14752},
  year = {2026},
}

Report

WALL-WM: Carving World Action Modeling at the Event Joints

Shalfun Li, Victor Yao, Charles Yang, …, Pushi Zhang, et al. 等 Shalfun Li, Victor Yao, Charles Yang, Truth Qu, Regis Cheng, Ryan Yu, Howard Lu, Newton Von, Vincent Chen, Yohann Tang, Maeve Zhang, Ellie Ma, Gody Li, Starrick Liu, Sage Yang, Lorien Shu, J. W. Gao, Ethan Chen, Colin Ye, Yu Sun, Elise Mon, Pushi Zhang, Neo Li, Lily Li, James Wang, Ping Yang, Chris Pan, Lucy Liang, Hang Su, Roy Gan, Hao Wang, and Qian Wang

Technical report, arXiv:2606.01955, 2026

Early version at ECCV 2026 Workshop (STEAI), Oral

arXiv Code Project

@article{li2026wallwm,
  title = {WALL-WM: Carving World Action Modeling at the Event Joints},
  author = {Li, Shalfun and Yao, Victor and Yang, Charles and Qu, Truth and Cheng, Regis and Yu, Ryan and Lu, Howard and Von, Newton and Chen, Vincent and Tang, Yohann and Zhang, Maeve and Ma, Ellie and Li, Gody and Liu, Starrick and Yang, Sage and Shu, Lorien and Gao, J. W. and Chen, Ethan and Ye, Colin and Sun, Yu and Mon, Elise and Zhang, Pushi and Li, Neo and Li, Lily and Wang, James and Yang, Ping and Pan, Chris and Liang, Lucy and Su, Hang and Gan, Roy and Wang, Hao and Wang, Qian},
  journal = {arXiv preprint arXiv:2606.01955},
  year = {2026},
  note = {Early version at ECCV 2026 Workshop (STEAI), Oral}
}

ICLR

villa-X: Enhancing Latent Action Modeling in Vision-Language-Action Models

Xiaoyu Chen^*, Hangxing Wei^*, Pushi Zhang^*, et al. 等 Xiaoyu Chen^*, Hangxing Wei^*, Pushi Zhang^*, Chuheng Zhang^*, Kaixin Wang^*, Yanjiang Guo, Rushuai Yang, Yucen Wang, Xinquan Xiao, Li Zhao^*, Jianyu Chen, and Jiang Bian

In International Conference on Learning Representations (ICLR), 2026

arXiv Code Model Project

@inproceedings{chen2026villax,
  title = {villa-X: Enhancing Latent Action Modeling in Vision-Language-Action Models},
  author = {Chen, Xiaoyu and Wei, Hangxing and Zhang, Pushi and Zhang, Chuheng and Wang, Kaixin and Guo, Yanjiang and Yang, Rushuai and Wang, Yucen and Xiao, Xinquan and Zhao, Li and Chen, Jianyu and Bian, Jiang},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year = {2026},
}

NeurIPS

What Do Latent Action Models Actually Learn?

Chuheng Zhang^*, Tim Pearce^*, Pushi Zhang, Kaixin Wang, Xiaoyu Chen, Wei Shen, Li Zhao, and Jiang Bian

In Advances in Neural Information Processing Systems (NeurIPS), 2025

arXiv

@inproceedings{zhang2025lam,
  title = {What Do Latent Action Models Actually Learn?},
  author = {Zhang, Chuheng and Pearce, Tim and Zhang, Pushi and Wang, Kaixin and Chen, Xiaoyu and Shen, Wei and Zhao, Li and Bian, Jiang},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  volume = {38},
  pages = {146676--146697},
  year = {2025},
}

Preprint

PIG-Nav: Key Insights for Pretrained Image Goal Navigation Models

Jiansong Wan, Chengming Zhou, Jinkua Liu, …, Pushi Zhang, et al. 等 Jiansong Wan, Chengming Zhou, Jinkua Liu, Xiangge Huang, Xiaoyu Chen, Xiaohan Yi, Qisen Yang, Baiting Zhu, Xin-Qiang Cai, Lixing Liu, Rushuai Yang, Chuheng Zhang, Sherif Abdelfattah, Hayong Shin, Pushi Zhang, Li Zhao, and Jiang Bian

Preprint, arXiv:2507.17220, 2025

arXiv Code Model Video

@article{wan2025pignav,
  title = {PIG-Nav: Key Insights for Pretrained Image Goal Navigation Models},
  author = {Wan, Jiansong and Zhou, Chengming and Liu, Jinkua and Huang, Xiangge and Chen, Xiaoyu and Yi, Xiaohan and Yang, Qisen and Zhu, Baiting and Cai, Xin-Qiang and Liu, Lixing and Yang, Rushuai and Zhang, Chuheng and Abdelfattah, Sherif and Shin, Hayong and Zhang, Pushi and Zhao, Li and Bian, Jiang},
  journal = {arXiv preprint arXiv:2507.17220},
  year = {2025},
}

Preprint

IGOR: Image-GOal Representations are the Atomic Control Units for Foundation Models in Embodied AI

Xiaoyu Chen^*, Junliang Guo^*, Tianyu He^*, Chuheng Zhang^*, Pushi Zhang, Derek Cathera Yang, Li Zhao^*, and Jiang Bian

Preprint, arXiv:2411.00785, 2024

arXiv Project

@article{chen2024igor,
  title = {IGOR: Image-GOal Representations are the Atomic Control Units for Foundation Models in Embodied AI},
  author = {Chen, Xiaoyu and Guo, Junliang and He, Tianyu and Zhang, Chuheng and Zhang, Pushi and Yang, Derek Cathera and Zhao, Li and Bian, Jiang},
  journal = {arXiv preprint arXiv:2411.00785},
  year = {2024},
}

NeurIPS

Distributional Pareto-Optimal Multi-Objective Reinforcement Learning

Xin-Qiang Cai^*, Pushi Zhang^*, Li Zhao, Jiang Bian, Masashi Sugiyama, and Ashley J. Llorens

In Advances in Neural Information Processing Systems (NeurIPS), 2023

PDF Code

@inproceedings{cai2023distributional,
  title = {Distributional Pareto-Optimal Multi-Objective Reinforcement Learning},
  author = {Cai, Xin-Qiang and Zhang, Pushi and Zhao, Li and Bian, Jiang and Sugiyama, Masashi and Llorens, Ashley J.},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  volume = {36},
  pages = {15593--15613},
  year = {2023},
}

NeurIPS

An Adaptive Deep RL Method for Non-Stationary Environments with Piecewise Stable Context

Xiaoyu Chen^*, Xiangming Zhu^*, Yufeng Zheng, Pushi Zhang, et al. 等 Xiaoyu Chen^*, Xiangming Zhu^*, Yufeng Zheng, Pushi Zhang, Li Zhao, Wenxue Cheng, Peng Cheng, Yongqiang Xiong, Tao Qin, Jianyu Chen, and Tie-Yan Liu

In Advances in Neural Information Processing Systems (NeurIPS), 2022

arXiv Code

@inproceedings{chen2022adaptive,
  title = {An Adaptive Deep RL Method for Non-Stationary Environments with Piecewise Stable Context},
  author = {Chen, Xiaoyu and Zhu, Xiangming and Zheng, Yufeng and Zhang, Pushi and Zhao, Li and Cheng, Wenxue and Cheng, Peng and Xiong, Yongqiang and Qin, Tao and Chen, Jianyu and Liu, Tie-Yan},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  volume = {35},
  pages = {35449--35461},
  year = {2022},
}

NeurIPS

Distributional Reinforcement Learning for Multi-Dimensional Reward Functions

Pushi Zhang^*, Xiaoyu Chen^*, Li Zhao, Wei Xiong, Tao Qin, and Tie-Yan Liu

In Advances in Neural Information Processing Systems (NeurIPS), 2021

arXiv PDF Code

@inproceedings{zhang2021distributional,
  title = {Distributional Reinforcement Learning for Multi-Dimensional Reward Functions},
  author = {Zhang, Pushi and Chen, Xiaoyu and Zhao, Li and Xiong, Wei and Qin, Tao and Liu, Tie-Yan},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  volume = {34},
  pages = {1519--1529},
  year = {2021},
}

Neurocomp.

Demonstration Actor Critic

Guoqing Liu, Li Zhao, Pushi Zhang, Jiang Bian, Tao Qin, Nenghai Yu, and Tie-Yan Liu

Neurocomputing, 2021

Paper

@article{liu2021demonstration,
  title = {Demonstration Actor Critic},
  author = {Liu, Guoqing and Zhao, Li and Zhang, Pushi and Bian, Jiang and Qin, Tao and Yu, Nenghai and Liu, Tie-Yan},
  journal = {Neurocomputing},
  volume = {434},
  pages = {194--202},
  year = {2021},
  publisher = {Elsevier},
  doi = {10.1016/j.neucom.2020.12.116},
}

Experience

2025.09 – Present Researcher and Project Lead, X Square Robot
2022.06 – 2025.09 Researcher, Machine Learning Group, Microsoft Research Asia

Education

2020.09 – 2022.06 M.S. in Computer and Information Science, University of Pennsylvania
2015.09 – 2019.06 B.S. in Computer Science, Tsinghua University