Changxun Pan

Undergraduate Researcher · Vision-Language-Action · Robot Learning

Changxun Pan

Junior undergraduate in the Yao Class, IIIS, Tsinghua University, and a research intern at Galaxea AI working with Prof. Hang Zhao.

I am interested in embodied agents that can reason and act over long temporal horizons. My current work focuses on memory-augmented Vision-Language-Action models, following earlier work on large-scale robot-data supervision for G0.5.

CV Email GitHub G0.5
01

About

I am a B.Eng. student in Computer Science and Technology in the Yao Class, Institute for Interdisciplinary Information Sciences (IIIS), Tsinghua University, expected to graduate in 2028.

Since January 2026, I have been a research intern at Galaxea AI, working directly with Prof. Hang Zhao. I first worked on scalable visual and geometric supervision for robot pretraining data, and I am now exploring memory mechanisms for long-horizon VLA policies.

02

Research

Current Research 2026 — Present

Memory-Augmented VLA Models for Long-Horizon Manipulation

I am exploring recurrent latent memory for embodied policies: using informative mid-layer hidden representations as compact memory that can be carried across policy inference steps while keeping training scalable.

  • Extract mid-layer hidden states, transform them through an adapter, and feed the resulting memory tokens into the next policy step.
  • Investigate an epoch-versioned memory bank that supports random-sample training without full backpropagation through time across long episodes.
  • Use G0.5 as the backbone and study memory-dependent manipulation with RMBench, MIKASA, and RoboMME; the project is still at an early experimental stage.
Selected Research Contribution · Galaxea AI Jan. 2026 — Present

G0.5 — Large-Scale Robot Data Supervision

As part of the G0.5 project, I focused on data and evaluation rather than model architecture, developing scalable object-centric and geometric supervision for robot pretraining data.

  • Built a two-stage bounding-box pipeline: generate task-relevant object boxes on keyframes with language-model prompting, then propagate them temporally with SAM 3; parallelized the stages across API workers and multiple GPUs.
  • Explored multimodal point-prediction and tracking approaches for 2D motion traces, then developed a faster URDF-based kinematic solution for three-view trajectories of dual-arm gripper/end-effector points on R1 Lite and R1 Pro.
  • The resulting supervision pipelines were applied to nearly 10,000 hours of robot data for pretraining. I directly validated their quality and reliability on several hundred hours; the auxiliary data also supported VQA-style VLM training.
Paper ↗
03

Publications

04

Honors & Education

Education

Tsinghua University
Yao Class, IIIS
B.Eng. in Computer Science and Technology
2024 — 2028
expected

Selected Coursework

Machine Learning · Natural Language Processing · Computer Vision · Advanced Computer Graphics · Large Language Model Applications · Multimodal Machine Learning

Honors

Chinese Physics Olympiad — National Gold Medal, ranked 25th nationally
National Encouragement Scholarship
Tsinghua University Comprehensive Excellence Scholarship

Skills

Programming: Python, C++
ML: PyTorch
Tools: Git, Linux