About

I am Jie He. My current research interests lie in vision-language-action models, embodied intelligence, robotic manipulation, and efficient model adaptation for real-world decision making.

News

1. Jun 2026CogVLA, H-GAR, and DeltaVLA have been added to the publication list.
2. Jun 2026Academic homepage moved to minijie3.github.io, with the blog hosted at /blog.

Research Interests

• Vision-Language-Action Models
• Embodied AI
• Robotic Manipulation
• Efficient Multimodal Learning

Publications

* Equal contribution. † Corresponding author.

ΔVLA: Prior-Guided Vision-Language-Action Models via World Knowledge Variation

Yijie Zhu, Jie He, Rui Shao†, Kaishen Yuan, Tao Tan, Xiaochen Yuan, Zitong Yu†

arXiv preprint arXiv:2603.08361, 2026.

ΔVLA studies how world knowledge variation can guide vision-language-action models. The work uses prior knowledge to improve action reasoning and adaptation, aiming to make embodied policies more reliable under changing task and scene conditions.

arXiv GitHub

H-GAR: A Hierarchical Interaction Framework via Goal-Driven Observation-Action Refinement for Robotic Manipulation

Yijie Zhu, Rui Shao†, Ziyang Liu, Jie He, Jizhihui Liu, Jiuru Wang, Zitong Yu†

AAAI 2026 (Oral).

H-GAR introduces a hierarchical interaction framework for robotic manipulation. It refines observations and actions according to task goals, improving the robot's ability to reason over long-horizon interactions and execute manipulation steps more precisely.

arXiv GitHub

CogVLA: Cognition-Aligned Vision-Language-Action Models via Instruction-Driven Routing and Sparsification

Wei Li, Renshan Zhang, Rui Shao†, Jie He, Liqiang Nie

NeurIPS 2025.

CogVLA aligns vision-language-action models with cognitive execution patterns through instruction-driven routing and sparsification. It targets efficient, task-aware reasoning so the model can activate the most relevant pathways for embodied decision making.

arXiv GitHub

Education