Released the training code, evaluation stack, and model artifacts for INTACT.
Ph.D. Student at Zhejiang University
Junhan Sun 孙俊涵
I am a first-year Ph.D. student in Computer Science at Zhejiang University, after completing my undergraduate study at Chu Kochen Honors College. I work with Prof. Guofeng Zhang and Prof. Hao Zhao on world models that connect predictive representations, latent dynamics, and efficient embodied control.
World Models
JEPA
Embodied AI
Robot Learning
College of Computer Science and Technology
State Key Laboratory of CAD&CG, Zhejiang University
Research premise
A useful world model should not stop at predicting what an action does. It should also learn which action realizes an intent.
Updates
News
Released the INTACT preprint and its interactive project page.
Open-sourced CLEAR-LeWM, an auditable evaluation suite for LeWM-compatible world models.
Research output
Publications
Selected work on predictive world representations and efficient embodied control.
arXiv · 2026 · Robotics
INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models
INTACT completes the intent-to-action loop in an end-to-end JEPA. It grounds deployable goal queries with local successor supervision, exposing a direct action interface while preserving the learned world representation.
- Training
- 1 epoch
- Direct macro SR
- 95.33%
- Test-time search
- 0
- Inference
- 2.9-5.5 ms
R1
Selected research software
CLEAR-LeWM
An auditable evaluation suite for LeWM-compatible latent world models, with official-compatible and difficulty-controlled task protocols.
Research
Questions I work on
I am interested in the geometry that makes predictive representations useful for physical reasoning and control.
Predictive representations
How should JEPA objectives and anti-collapse regularization shape a latent space that remains stable, informative, and easy to model?
Intent-conditioned control
How can action supervision turn latent change into a deployable control interface without sacrificing the world model's representation quality?
Efficient embodied learning
How can one compact model learn across tasks, train with limited compute, and replace expensive online search with responsive closed-loop inference?
Journey
Education
2026 - Present
Ph.D. Student in Computer Science
Zhejiang University
College of Computer Science and Technology. Researching JEPA world models, representation learning, and embodied control.
2022 - 2026
Undergraduate Program
Chu Kochen Honors College, Zhejiang University
Built a foundation across computer science, visual computing, machine learning, and research practice before continuing directly to the Ph.D. program.
Contact
Let us compare ideas, results, and assumptions.
I welcome conversations about world models, JEPA, representation geometry, robot learning, reproducibility, and research collaboration.