Robot Learning · VLA / WAM · Continual Learning
From robot data to real-world deployment.
I'm Siyuan Luo, a Master of Software Engineering student at the University of Melbourne. I build embodied-AI systems across data collection, model training, policy serving, and real-robot execution - recently π₀.₅ on Franka and whole-body world-action models on Unitree G1.
Current focus
Whole-Body WAM
Extending FastWAM from tabletop manipulation to whole-body loco-manipulation on Unitree G1.
Project details →π₀.₅ on Franka
An end-to-end VLA pipeline spanning DROID-style data, LeRobot, full fine-tuning, serving, and real-robot execution.
Project details →Audio-L3A
Closed-form audio class-incremental learning that approaches the offline upper bound at a fraction of the training cost.
Project details →Latest
- 2026-08 - Built a Kubernetes-based remote review workbench for multi-source whole-body robot datasets.
- 2026-05 - Joined the Xiong'an Institute of Artificial Intelligence as an algorithm intern, working on Whole-Body WAM for Unitree G1.
- 2026-05 - Audio-L3A reached 45.16% mAP on five-phase AudioSet-50, 2.65 points above the strongest CIL baseline.
- 2026-04 - Completed the π₀.₅ training-to-deployment pipeline for Franka flexible manipulation.
- 2026-01 - Completed an AI application internship at KNQ Technology, delivering an end-to-end table-tennis commentary system.
Experience
Algorithm Intern, Xiong’an Institute of Artificial Intelligence · 2026.05 - 2026.08
Working on FastWAM-based whole-body loco-manipulation for Unitree G1: continuous 72-D physical actions, joint video-action flow matching, multi-source data alignment, distributed training, VR teleoperation, and demonstration-data quality control.
AI Application Engineer Intern, KNQ Technology · 2025.11 - 2026.01
Built a real-time table-tennis commentary pipeline in Python and FastAPI: video input → OCR → LLM generation → TTS → video/RTMP output. Added structured match memory, pronunciation correction, commentary modes, and a pre-match narrative agent.
Education
Master of Software Engineering, University of Melbourne · 2024.07 - present
Supervised by Dr. Ting Dang. Research in audio class-incremental learning.
Bachelor of Engineering in Cyberspace Security, Sichuan University · 2020.09 - 2024.06
Get in touch
Email luosylois@gmail.com · GitHub @LUOSYrrrr · CV (PDF)
机器人学习 · VLA / WAM · 持续学习
把机器人学习从数据做到真机。
我是骆思缘，墨尔本大学软件工程硕士生。我关注具身智能系统的完整链路：示范数据采集、模型训练、策略服务与真机执行。近期工作包括 Franka 上的 π₀.₅，以及面向 Unitree G1 的全身 World-Action Model。
当前重点
Whole-Body WAM
将 FastWAM 从桌面机械臂任务扩展到 Unitree G1 全身移动操作。
项目详情 →Franka 上的 π₀.₅
覆盖 DROID 风格数采、LeRobot 数据、全量微调、策略服务与真机执行的 VLA 全链路。
项目详情 →Audio-L3A
面向音频类增量学习的闭式解方法，以更低训练成本逼近离线训练上界。
项目详情 →最新进展
- 2026-08 - 完成面向多源全身机器人数据的 Kubernetes 远程质检与人工审核工作台。
- 2026-05 - 加入雄安人工智能研究院担任算法实习生，参与 Unitree G1 Whole-Body WAM。
- 2026-05 - Audio-L3A 在五阶段 AudioSet-50 上达到 45.16% mAP，领先最强 CIL 基线 2.65 个百分点。
- 2026-04 - 打通 Franka 柔性操作场景下 π₀.₅ 从训练到真机部署的完整链路。
- 2026-01 - 完成麒纪科技 AI 应用开发实习，交付乒乓球赛事 AI 实时解说系统。
工作经历
雄安人工智能研究院，算法实习生 · 2026.05 - 2026.08
参与基于 FastWAM 的 Unitree G1 全身移动操作研究，涉及 72 维连续物理动作、视频-动作联合 Flow Matching、多源数据统一、分布式训练、VR 遥操作和示范数据质检。
北京麒纪科技，AI 应用开发工程师（实习） · 2025.11 - 2026.01
使用 Python + FastAPI 打通视频输入 → OCR → LLM 生成 → TTS → 视频/RTMP 输出的实时解说链路，并实现赛事记忆、读音修正、多档解说模式和预赛备稿 Agent。
教育背景
墨尔本大学，软件工程硕士 · 2024.07 - 至今
导师：Dr. Ting Dang。研究方向：音频类增量学习。
四川大学，网络空间安全学士 · 2020.09 - 2024.06
联系方式
邮箱 luosylois@gmail.com · GitHub @LUOSYrrrr · 简历（PDF）