Zhuoyang Wu

吴卓阳Zhuoyang Wu

Zhuoyang WuSYSU · AI

机械义肢主题的 Q 版人物贴纸

探索具身智能和 AI Agent，目前主要在机械臂上实践 VLA 模型，并关注世界模型。

I explore embodied intelligence andAI agents, currently experimenting with VLA models on robot arms and following work on world models.

GoPythonC++PyTorchROSLangGraphPostgreSQLRedisDocker
GitHub服务器ServersEmail
微信WeChat
woshiwudachong
↓ scroll

关于About

我是中山大学网络空间安全学院本科生（2027 届），就读于深圳校区，GPA 3.8/4。

过去约一年半主要参与大模型训练数据去重相关研究，相关论文分别在投 DCC 和 WWW。目前在灵巧智能 DEXGEM Lab 担任算法工程师（实习）。此前在深信服科技实习，负责自进化项目的数据挖掘算法开发。

最近在做机械臂和仿真相关的实践，包括在 SO-101 上运行 ACT、π0.5，以及用 Isaac Lab 搭建抓放环境。我还在学习和摸索，希望通过具体项目，把论文里的方法、代码实现和实际运行中的问题逐渐联系起来。这个博客主要记录过程、遇到的问题，以及目前的理解。

"做一点有意义的工作"

I'm an undergraduate at Sun Yat-sen University's School of Cyber Science and Technology, based on the Shenzhen campus and graduating in 2027. My GPA is 3.8/4.

Over roughly the past year and a half, I've mainly worked on training-data deduplication for large language models. Papers from this work are under review at DCC and WWW. I'm currently an algorithm engineer intern at DEXGEM Lab. Previously, I interned at Sangfor Technologies, developing data-mining algorithms for a self-evolving system.

Recently, I've been working on robot arms and simulation: running ACT and π0.5 on SO-101 and building a pick-and-place environment in Isaac Lab. I'm still learning and figuring things out, using concrete projects to connect methods from papers with their code and the problems that arise when running them. This blog records that process, the issues I encounter, and my current understanding.

"Do work that matters."

项目Projects

MMdedupMMdedup: Multimodal Deduplication Is More Than Similarity Search

一个面向文本、图像、音频及图文对的数据清洗研究项目，重点区分候选召回、重复关系判断与安全删除，并为实验结果保留可追溯证据。A research system for cleaning text, image, audio, and image-caption data, separating candidate retrieval, duplicate relation typing, and safe deletion with traceable experiment evidence.

PythonPython多模态Multimodal数据去重Data Deduplication实验系统Experiment Systems研究Research

深信服实习Sangfor Internship: Reducing 10,000 Security Alerts to 680 Review Candidates

为终端安全数据构建语义与行为图双维度的增量聚类管道，处理生长告警、簇心更新和运营 review 包交付。An incremental clustering pipeline for endpoint-security data, combining semantic and behavior-graph signals with growth-event deduplication and review-package delivery.

PythonPython聚类Clustering图算法Graph Algorithms安全数据Security Data实习Internship

yu_agentyu_agent: Rebuilding the Core Pieces to Understand Agents

一个学习阶段的 Python Agent 框架：实现统一 LLM 接口、四类 Agent、工具与 MCP、分层记忆和 RAG，也诚实记录尚未解决的路由与自修正问题。A learning-stage Python agent framework implementing a unified LLM interface, four agent types, tools and MCP, layered memory, and RAG—along with its unresolved routing and self-correction limits.

PythonPythonAgentAgentRAGRAGMCPMCPQdrantQdrantNeo4jNeo4j

JunkFilterJunkFilter: Letting an Agent Read the RSS Backlog First

一个由 Go、Python、Redis Stream 和 PostgreSQL 组成的个人信息过滤系统：抓取 RSS，去重并评估内容，再通过网页和 Telegram 交互。A personal information-filtering system built with Go, Python, Redis Streams, and PostgreSQL for RSS ingestion, deduplication, LLM evaluation, and web or Telegram interaction.

GoGoPythonPythonLangGraphLangGraphRedis StreamRedis StreamsAgentAgent

写作Writing

XPolicyLabFirst Notes on XPolicyLab

从 π0.5 的输入输出适配，看 XPolicyLab 怎样复用模型接入代码，连接仿真与真机。How XPolicyLab adapts π0.5 inputs and outputs and reuses policy integrations across simulation and real robots.

2026-10-05

LingBot-VLA 2.0 初读First Notes on LingBot-VLA 2.0

梳理 LingBot-VLA 2.0 的数据对齐、55 维动作空间、MoE、流匹配与未来时空监督。Reading LingBot-VLA 2.0 through data alignment, its 55-dimensional action space, sparse MoE, flow matching, and future-aware distillation.

2026-09-22

Isaac Lab 抓放Building a Four-Tool Pick-and-Place Task in Isaac Lab

从 Franka 键盘控制、三相机录像到成功判据与 HDF5 回放，记录一次可安装仿真环境的实现与验证。An implementation record covering Franka keyboard control, three-camera demonstrations, success metrics, and HDF5 replay.

2026-09-17

π0.5 初读First Notes on π0.5

结合论文与阅读讨论，梳理 π0.5 的异构数据、高低层架构、FAST、流匹配和联合损失。Reading notes on π0.5: heterogeneous co-training, hierarchical inference, FAST, flow matching, and the joint objective.

2026-09-13

ACT 原文阅读Reading ACT: From Action Chunks to Our Final 50-Demonstration Retraining

把 ACT 论文的低成本 ALOHA、Action Chunking、CVAE 和 Temporal Ensembling，和我们在 SO-101 上最后一次 50 条示教重训放在一起看。A plain-language reading of ALOHA, Action Chunking, CVAE, and Temporal Ensembling, connected to our final 50-demonstration ACT retraining on SO-101.

2026-09-12

SO-101 仿真Reproducing the SO-101 Tape Task in Isaac Sim

从真机尺寸、腕部视角和版本隔离开始，记录 SO-101 蓝色胶带任务的数字孪生搭建，以及一次没有被包装成成功的抓取诊断。Building a digital twin of the SO-101 tape task from real measurements and wrist views, then diagnosing a grasp that still failed the physics gate.

2026-08-16

SO-101 上的 π0.5Running π0.5 on a Real SO-101: Remote Inference and Action Handoff

记录从 50 条真机示教、RTX 3090 远程推理，到三种动作交接路线对比的 π0.5 真机闭环。A practical record of 50 real-robot demonstrations, remote π0.5 inference on an RTX 3090, and three action-handoff strategies.

2026-08-15

SO-101 上的 ACTReproducing ACT on SO-101 with 9 Demonstrations

记录我从采集第一批示教数据，到训练 ACT、排查动作块问题，最终让 SO-101 在固定场景完成抓放任务的过程。A practical record of collecting demonstrations, training ACT, debugging action chunks, and closing the loop on a real SO-101 robot.

2026-08-13

SO-101 组装From a Desk Full of Parts to a Working SO-101

记录 SO-101 主从机械臂从舵机准备、结构装配到上电检查的过程，以及几个最容易返工的地方。A hands-on record of preparing the servos, assembling the leader and follower arms, and checking the hardware before power-up.

2026-08-10