Zhirui Zhang / DeepCybo
Partner & Head of Physical Foundation Models
I lead DeepCybo’s physical foundation model program, bringing perception, reasoning, and action into the real world.
01 / Selected work
Current research at DeepCybo.
From foundation models to robotic action.
One model. A shared token space.
01 — Physical foundation models
Embodied understanding, action generation, and future-state prediction in one autoregressive model.
02 — Human learning, robot action
Turning synchronized human videos into 60-DoF action supervision for real humanoid manipulation.
Cup stacking, ring toss, and water pouring without target-task robot demonstrations.
02 / About
Zhirui ZhangResearcher. Builder. Team leader.
I lead the technical roadmap and R&D for physical foundation models at DeepCybo, spanning models, data, training, and evaluation.
My work connects large-scale pretraining, post-training, and deployed AI systems. I received my Ph.D. from USTC through a joint training program with Microsoft Research Asia.
03 / Experience
Partner · Head of Physical Foundation Models
Leading the physical foundation model roadmap, team, and R&D.
AI Technical Advisor · MoonBit model-and-tool systems and SWE-AGI.
Technical Expert · Led Xiaoyi post-training, reasoning, and model deployment.
Algorithm Expert · Step2 MoE pretraining, FP8, and long-context research.
Senior Researcher · Translation platforms and multilingual models.
Algorithm Expert · Multilingual translation research and commercial systems.
Research Intern · Neural machine translation and language generation.
04 / Publications
Technical Report 2018
EMNLP 2023
ACL-IJCNLP 2021
AAAI 2019
NeurIPS 2020
ICLR 2023
arXiv 2026 · Paper ↗
arXiv 2026 · Paper ↗
arXiv 2026 · Paper ↗
arXiv 2026 · Paper ↗
arXiv 2026 · Paper ↗
ACL 2026 · Paper ↗
05 / Beyond the papers
On the capability boundaries of foundation models.
Research / Ideas / Collaboration