Changxun Pan
Yao Class, Institute for Interdisciplinary Information Sciences (IIIS), Tsinghua University
Research Interests
Vision-Language-Action Models, Robot Learning, Memory for Embodied Agents, Multimodal Learning
Education
Tsinghua University, Beijing
B.Eng. in Computer Science and Technology, Yao Class, IIIS
B.Eng. in Computer Science and Technology, Yao Class, IIIS
Sept. 2024 — Jul. 2028 (expected)
Research Experience
Galaxea AI — Research Intern
Advisor: Prof. Hang Zhao
Advisor: Prof. Hang Zhao
Jan. 2026 — Present
- Contributed to large-scale object-centric and geometric supervision for G0.5 robot pretraining data.
- Built a two-stage bounding-box pipeline using language-model prompting for task-relevant keyframes and SAM 3 for temporal propagation.
- Parallelized language-model inference across workers and mask propagation across multiple GPUs for efficient scaling.
- Investigated multimodal point-prediction and tracking approaches for 2D motion traces, then developed a faster URDF-based solution.
- Generated three-view 2D trajectories of dual-arm gripper/end-effector points for R1 Lite and R1 Pro using kinematic projection and GPU acceleration.
- The pipelines were applied to nearly 10,000 hours of robot data for pretraining; directly validated quality and reliability on several hundred hours.
Memory-Augmented VLA Models for Long-Horizon Manipulation
2026 — Present
- Exploring recurrent mid-layer latent memory, where adapted hidden representations are carried forward as memory tokens for the next policy inference.
- Investigating a memory-bank training scheme that supports random sampling while avoiding full BPTT across long episodes.
- Current experiments use G0.5 as the backbone and memory-oriented benchmarks including RMBench, MIKASA, and RoboMME.
Preprint
G0.5: One Autoregressive Stream for Robot Reasoning and Action
Yicheng Liu, Zibin Dong, Baijun Ye, …, Changxun Pan, …, Hang Zhao
arXiv preprint, 2026
Yicheng Liu, Zibin Dong, Baijun Ye, …, Changxun Pan, …, Hang Zhao
arXiv preprint, 2026
Honors & Awards
Chinese Physics Olympiad — National Gold Medal, ranked 25th nationally2023
National Encouragement Scholarship2025
Tsinghua University Comprehensive Excellence Scholarship2025
Selected Coursework & Skills
Coursework: Machine Learning, Natural Language Processing, Computer Vision, Advanced Computer Graphics, Large Language Model Applications, Multimodal Machine Learning
Programming: Python, C++ · ML: PyTorch · Tools: Git, Linux