[{"data":1,"prerenderedAt":346},["ShallowReactive",2],{"project-detail-\u002Fprojects\u002Flerobot-vla-reproduction":3,"project-detail-nav":100},{"id":4,"title":5,"body":6,"category":56,"cover":57,"description":58,"extension":59,"highlights":60,"images":65,"links":66,"meta":68,"navigation":70,"path":71,"period":72,"portrait":73,"poster":57,"role":74,"seo":75,"stack":76,"status":81,"stem":82,"tags":83,"video":87,"videos":88,"__hash__":99},"projects\u002Fprojects\u002Flerobot-vla-reproduction.md","LeRobot101：ACT \u002F smolVLA \u002F pi0.5 复现",{"type":7,"value":8,"toc":49},"minimark",[9,13,17,20,43,46],[10,11,12],"h2",{"id":12},"项目简介",[14,15,16],"p",{},"以 Hugging Face LeRobot 框架为主线，系统复现 ACT、smolVLA、pi0.5 三个代表性模型，覆盖模仿学习与视觉-语言-动作两大技术路线。项目打通了从遥操作数据采集、策略训练到仿真评估的完整链路，所有复现代码已整理上传至 GitHub。",[10,18,19],{"id":19},"复现内容",[21,22,23,31,37],"ul",{},[24,25,26,30],"li",{},[27,28,29],"strong",{},"ACT（Action Chunking with Transformers）","：以动作分块预测替代单步动作回归，配合时间集成（temporal ensembling）平滑执行轨迹，复现数据采集、训练与仿真评估流程。",[24,32,33,36],{},[27,34,35],{},"smolVLA","：基于 SmolVLM 视觉-语言骨干的轻量 VLA 模型，在桌面操作任务上完成端到端复现，理解视觉观测与语言指令到动作 token 的映射方式。",[24,38,39,42],{},[27,40,41],{},"pi0.5","：基于 OpenPI 的预训练 VLA 策略，在自有任务配置上进行微调，验证「语言指令 → 视觉-语言理解 → 动作序列输出」的完整执行链路。",[10,44,45],{"id":45},"评估结果",[14,47,48],{},"在 Robosuite 中以脚本专家生成演示数据，训练行为克隆（BC）策略；在连续 20 轮评估中达到 80% 任务成功率，验证了数据质量与动作分块对策略稳定性的作用。",{"title":50,"searchDepth":51,"depth":51,"links":52},"",2,[53,54,55],{"id":12,"depth":51,"text":12},{"id":19,"depth":51,"text":19},{"id":45,"depth":51,"text":45},"机器人学习","\u002Fimages\u002Fprojects\u002Fact-poster.jpg","基于 LeRobot 框架复现 ACT、smolVLA 与 pi0.5 三类模仿学习与视觉-语言-动作模型，打通数据采集、策略训练到仿真评估全流程。","md",[61,62,63,64],"跟随 LeRobot101 复现 ACT 模型，理解动作分块（Action Chunking）与模仿学习训练流程。","复现 smolVLA 轻量 VLA 模型，在仿真桌面任务上完成数据采集、策略训练与评估。","复现 pi0.5：基于 OpenPI 预训练 VLA 策略进行微调，验证语言指令到机械臂动作的执行链路。","配合 Robosuite 行为克隆实验，从脚本专家数据训练策略，20 轮评估达到 80% 成功率。",null,{"github":67},"https:\u002F\u002Fgithub.com\u002Fwkcherish\u002Flerobot.git",{"date":69},"2026-05-20",true,"\u002Fprojects\u002Flerobot-vla-reproduction","2025.11 – 2026.05",false,"独立复现",{"title":5,"description":58},[77,78,79,80],"LeRobot","PyTorch","Transformers","MuJoCo","已完成","projects\u002Flerobot-vla-reproduction",[77,84,35,41,85,86],"ACT","模仿学习","VLA","\u002Fvideos\u002Fprojects\u002Fact.mp4",[89,91,95],{"src":87,"title":90,"poster":57},"ACT 策略仿真任务演示",{"src":92,"title":93,"poster":94},"\u002Fvideos\u002Fprojects\u002Fsmolvla.mp4","smolVLA 桌面任务复现","\u002Fimages\u002Fprojects\u002Fsmolvla-poster.jpg",{"src":96,"title":97,"poster":98},"\u002Fvideos\u002Fprojects\u002Fpi0.5.mp4","pi0.5 微调策略执行","\u002Fimages\u002Fprojects\u002Fpi05-poster.jpg","OBkLJ9rOpfpt66cMl1TkSBaef3E8ALjvOwncZv3t9WU",[101,179,267,309],{"id":102,"title":103,"body":104,"category":145,"cover":146,"description":147,"extension":59,"highlights":148,"images":153,"links":159,"meta":161,"navigation":70,"path":163,"period":164,"portrait":73,"poster":65,"role":165,"seo":166,"stack":167,"status":172,"stem":173,"tags":174,"video":177,"videos":65,"__hash__":178},"projects\u002Fprojects\u002Far-ur5-real.md","真机拓展：Quest 3 AR 透视控制真实 UR5",{"type":7,"value":105,"toc":140},[106,108,111,114,134,137],[10,107,12],{"id":12},[14,109,110],{},"在「Quest 3 VR 远程操控 UR5 数字孪生」纯仿真项目的基础上，独立开发真机控制链路 UR5_Teleop：Quest 3 以 AR 彩色透视模式运行，将头显与手柄位姿映射到真实 UR5 机械臂，在实际工作台上完成遥控运动。该拓展与仿真仓库代码相互独立。",[10,112,113],{"id":113},"技术方案",[21,115,116,122,128],{},[24,117,118,121],{},[27,119,120],{},"真机通信","：头显控制位姿经局域网发送至真机控制服务，转换为 UR5 可执行的末端目标。",[24,123,124,127],{},[27,125,126],{},"AR 透视","：利用 Quest 3 彩色透视直接观察真实工作台与机械臂，无需重建场景模型。",[24,129,130,133],{},[27,131,132],{},"控制迁移","：复用仿真中验证的笛卡尔伺服与夹爪连续控制逻辑，缩短真机调试周期。",[10,135,136],{"id":136},"个人收获",[14,138,139],{},"完整走通了控制策略从仿真到真机的迁移过程，体会到 AR 透视在遥操作中降低空间认知负担的作用，也积累了真机安全控制与通信时延处理的经验。",{"title":50,"searchDepth":51,"depth":51,"links":141},[142,143,144],{"id":12,"depth":51,"text":12},{"id":113,"depth":51,"text":113},{"id":136,"depth":51,"text":136},"具身智能","\u002Fimages\u002Fprojects\u002Freal-hardware-overview.jpg","在纯仿真项目基础上独立拓展真机控制链路 UR5_Teleop，Quest 3 AR 彩色透视下将头显与手柄位姿映射到真实 UR5 完成遥控运动。",[149,150,151,152],"在纯仿真项目基础上独立拓展真机控制链路（UR5_Teleop），将 Quest 3 头显与手柄位姿映射到真实 UR5。","Quest 3 AR 彩色透视下保留真实工作台视野，在实际工作台上完成机械臂遥控运动。","真机端复用笛卡尔伺服与连续夹爪控制逻辑，仿真中验证的控制链路可直接向真机迁移。","拓展代码 UR5_Teleop 与仿真仓库相互独立，已开源。",[154,156],{"src":146,"title":155},"真机 UR5、RobotiQ 夹爪与 D405 腕部相机",{"src":157,"title":158},"\u002Fimages\u002Fprojects\u002Freal-d405-mount.jpg","D405 相机法兰安装",{"github":160},"https:\u002F\u002Fgithub.com\u002Fwkcherish\u002FUR5_Teleop.git",{"date":162},"2026-09-01","\u002Fprojects\u002Far-ur5-real","2026.08 – 至今","独立开发",{"title":103,"description":147},[168,169,170,171],"Quest 3","Python","UR5","Three.js","进行中","projects\u002Far-ur5-real",[168,126,170,175,176],"UR5_Teleop","遥操作","\u002Fvideos\u002Fprojects\u002Far-ur5.mp4","Piwg_HpfiihFgsBjLUZdewuitikfE9K5-oasSLhNyGM",{"id":180,"title":181,"body":182,"category":145,"cover":233,"description":234,"extension":59,"highlights":235,"images":240,"links":255,"meta":256,"navigation":70,"path":258,"period":259,"portrait":73,"poster":65,"role":165,"seo":260,"stack":261,"status":81,"stem":263,"tags":264,"video":50,"videos":65,"__hash__":266},"projects\u002Fprojects\u002Far-ur5-rl.md","Quest 3 VR 远程操控 UR5 数字孪生（纯仿真）",{"type":7,"value":183,"toc":228},[184,186,189,191,223,225],[10,185,12],{"id":12},[14,187,188],{},"搭建可用于交互控制和状态反馈的 UR5 机械臂 MuJoCo 数字孪生仿真。前端基于 Three.js 构建与后端状态相互映射的 3D 可视化界面，后端使用 FastAPI 接收控制输入并驱动仿真；Quest 3 通过局域网 WebXR 进入 VR 后，可用手柄或手势连续控制机械臂与夹爪。本项目为纯仿真工作，真机 AR 拓展为独立项目。",[10,190,113],{"id":113},[21,192,193,199,205,211,217],{},[24,194,195,198],{},[27,196,197],{},"仿真环境","：MuJoCo 中建立 UR5 机械臂、工作台、工具箱与工具模型，配置接触、摩擦与执行器参数。",[24,200,201,204],{},[27,202,203],{},"可视化","：Three.js 实时渲染机械臂状态，与后端通过 WebSocket\u002FHTTP 保持状态同步。",[24,206,207,210],{},[27,208,209],{},"通信层","：FastAPI 提供状态观测、相机画面与控制接口，桥接前端输入与仿真动作。",[24,212,213,216],{},[27,214,215],{},"VR 交互","：Quest 3 利用手柄 6DoF 位姿或 Hand Tracking 手势，经笛卡尔伺服映射为末端执行器目标位姿。",[24,218,219,222],{},[27,220,221],{},"伺服控制","：后端通过 MuJoCo Jacobian 将末端误差换算为关节速度，并串联限位、奇异点降速与碰撞保护。",[10,224,136],{"id":136},[14,226,227],{},"系统理解了数字孪生系统从建模、通信到 VR 交互的完整链路，掌握了 WebXR 开发与基于 Jacobian 的笛卡尔伺服方法，也为后续真机遥操作与模仿学习复现打下基础。",{"title":50,"searchDepth":51,"depth":51,"links":229},[230,231,232],{"id":12,"depth":51,"text":12},{"id":113,"depth":51,"text":113},{"id":136,"depth":51,"text":136},"\u002Fimages\u002Fprojects\u002Fsim-overview.png","基于 MuJoCo 的 UR5 数字孪生仿真，Three.js 可视化与 FastAPI 后端，Quest 3 通过 WebXR 在 VR 中以手柄或手势控制机械臂。",[236,237,238,239],"在 MuJoCo 中搭建 UR5 机械臂、工作台与工具箱、钳子、扳手等工具的高保真仿真场景，FastAPI 后端提供状态观测与控制接口。","Three.js 构建与后端状态实时双向映射的 3D 数字孪生界面，桌面浏览器即可完成调试与观测。","Quest 3 经局域网以 WebXR 进入 VR，手柄或手势通过 6DoF 笛卡尔伺服控制机械臂，Trigger 连续控制夹爪开合。","后端基于 Jacobian 实现 Cartesian servo，并施加速度、加速度、关节限位与奇异点约束，保证仿真中运动平滑安全。",[241,243,246,249,252],{"src":233,"title":242},"MuJoCo 仿真场景：UR5、工作台与工具箱",{"src":244,"title":245},"\u002Fimages\u002Fprojects\u002Fsim-overview-2.png","仿真场景另一视角",{"src":247,"title":248},"\u002Fimages\u002Fprojects\u002Fsim-overview-3.png","工作台后侧视角",{"src":250,"title":251},"\u002Fimages\u002Fprojects\u002Fsim-d405-view.png","D405 腕部相机渲染视角",{"src":253,"title":254},"\u002Fimages\u002Fprojects\u002Fsim-cad-mount.png","相机法兰安装件设计",{},{"date":257},"2026-06-15","\u002Fprojects\u002Far-ur5-rl","2026.03 – 2026.06",{"title":181,"description":234},[80,171,262,168,169],"FastAPI","projects\u002Far-ur5-rl",[80,171,262,168,265],"WebXR","kXCTEg8RB2d0GP8OHN0IfhCQg71WeWqr28NSnqb6jHU",{"id":4,"title":5,"body":268,"category":56,"cover":57,"description":58,"extension":59,"highlights":299,"images":65,"links":300,"meta":301,"navigation":70,"path":71,"period":72,"portrait":73,"poster":57,"role":74,"seo":302,"stack":303,"status":81,"stem":82,"tags":304,"video":87,"videos":305,"__hash__":99},{"type":7,"value":269,"toc":294},[270,272,274,276,290,292],[10,271,12],{"id":12},[14,273,16],{},[10,275,19],{"id":19},[21,277,278,282,286],{},[24,279,280,30],{},[27,281,29],{},[24,283,284,36],{},[27,285,35],{},[24,287,288,42],{},[27,289,41],{},[10,291,45],{"id":45},[14,293,48],{},{"title":50,"searchDepth":51,"depth":51,"links":295},[296,297,298],{"id":12,"depth":51,"text":12},{"id":19,"depth":51,"text":19},{"id":45,"depth":51,"text":45},[61,62,63,64],{"github":67},{"date":69},{"title":5,"description":58},[77,78,79,80],[77,84,35,41,85,86],[306,307,308],{"src":87,"title":90,"poster":57},{"src":92,"title":93,"poster":94},{"src":96,"title":97,"poster":98},{"id":310,"title":311,"body":312,"category":322,"cover":323,"description":324,"extension":59,"highlights":325,"images":65,"links":330,"meta":331,"navigation":70,"path":333,"period":334,"portrait":73,"poster":65,"role":165,"seo":335,"stack":336,"status":172,"stem":340,"tags":341,"video":50,"videos":65,"__hash__":345},"projects\u002Fprojects\u002Furgrasp-agent.md","UR5 Grasp Agent：从手眼标定到 GraspNet 抓取",{"type":7,"value":313,"toc":319},[314,316],[10,315,12],{"id":12},[14,317,318],{},"构建一套从硬件标定到抓取策略的 UR5 抓取系统，打通 RealSense 感知、GraspNet 位姿检测与机械臂执行链路。",{"title":50,"searchDepth":51,"depth":51,"links":320},[321],{"id":12,"depth":51,"text":12},"机器人感知","\u002Fimages\u002Fprojects\u002Furgrasp-annotation.png","面向真实 UR5 的抓取智能体，覆盖相机标定、手眼标定、GraspNet 抓取位姿检测与机械臂运动执行全链路。",[326,327,328,329],"完成 RealSense 相机内参标定与眼在手上（eye-in-hand）手眼标定，统一米制与旋转矢量表示。","基于 GraspNet 从 RGB-D 场景检测 6D 抓取位姿，融合桌面物体分割与抓取点生成。","通过 ur_rtde 驱动 UR5 执行抓取运动，构建从视觉感知到运动执行的完整抓取 Agent。","统一相机系、标定板系与机器人基坐标系的米制尺度与旋转矢量表示，控制手眼标定重投影与位姿传递误差。",{},{"date":332},"2025-09-01","\u002Fprojects\u002Furgrasp-agent","2025.12 – 2026.04",{"title":311,"description":324},[169,78,337,338,339],"OpenCV","GraspNet","ur_rtde","projects\u002Furgrasp-agent",[338,170,342,343,344],"手眼标定","RealSense","6D 抓取位姿","YyQkluZioC8fIDzM_lYdoYv_sstvJ1FHw1GVcy43mYU",1791660993167]