[{"data":1,"prerenderedAt":100},["ShallowReactive",2],{"\u002Fglossary\u002Fsim-to-real":3},{"id":4,"title":5,"alternateName":6,"body":7,"description":90,"extension":91,"keywords":92,"meta":93,"navigation":94,"path":95,"seo":96,"stem":97,"updated":98,"__hash__":99},"glossary\u002Fglossary\u002Fzh\u002Fsim-to-real.md","Sim-to-Real（仿真到实机）","仿真到实机迁移",{"type":8,"value":9,"toc":84},"minimark",[10,15,23,28,35,63,66],[11,12,14],"h1",{"id":13},"sim-to-real-是什么","Sim-to-Real 是什么？",[16,17,18,22],"p",{},[19,20,21],"strong",{},"Sim-to-Real","（仿真到实机）是指先在物理仿真器（如 MuJoCo、Isaac Gym\u002FLab）中训练机器人控制策略，再把训练好的策略部署到真实机器人上的技术路线。真机训练又慢又贵还会摔坏硬件，而仿真可以并行跑成千上万个机器人、以超实时速度积累\"经验\"，这让强化学习训练人形机器人行走等技能成为可能。",[24,25,27],"h2",{"id":26},"核心挑战现实差距","核心挑战：现实差距",[16,29,30,31,34],{},"仿真永远不完全等于现实（摩擦、延迟、电机特性、传感噪声的差异），这个差异称为 ",[19,32,33],{},"Sim-to-Real Gap（现实差距）","。主流缓解手段：",[36,37,38,45,51],"ul",{},[39,40,41,44],"li",{},[19,42,43],{},"域随机化","：训练时随机扰动仿真参数（质量、摩擦、延迟），逼策略学会鲁棒性；",[39,46,47,50],{},[19,48,49],{},"精确系统辨识","：把真机的电机响应、惯量参数测准后写回仿真；",[39,52,53,56,57,62],{},[19,54,55],{},"执行器建模","：对",[58,59,61],"a",{"href":60},"\u002Fglossary\u002Fjoint-motor","关节电机","的扭矩-速度特性单独建模，这是腿式机器人迁移成败的关键之一。",[24,64,65],{"id":65},"对硬件的要求",[16,67,68,69,73,74,78,79,83],{},"策略在实机上以数百 Hz 频率输出关节指令，要求执行器具备高带宽力控（见",[58,70,72],{"href":71},"\u002Fglossary\u002Fquasi-direct-drive","准直驱","与 ",[58,75,77],{"href":76},"\u002Fglossary\u002Fmit-protocol-can","MIT 协议","）与低时延总线。BXI ",[58,80,82],{"href":81},"\u002Frobots\u002Fhumanoid-robot","Elf 3 人形机器人","内置 MuJoCo 仿真环境与 ROS2 SDK，配合 >1000 Hz 的 PCIE-CANFD 控制架构，支持从仿真到实机的完整工作流。",{"title":85,"searchDepth":86,"depth":86,"links":87},"",2,[88,89],{"id":26,"depth":86,"text":27},{"id":65,"depth":86,"text":65},"Sim-to-Real 指先在物理仿真环境中大规模训练机器人控制策略，再迁移到真实机器人上运行的技术路线，是当前人形机器人运动控制的主流训练范式。","md","sim-to-real, 仿真到实机, 强化学习机器人, 域随机化, MuJoCo, Isaac",{},true,"\u002Fglossary\u002Fzh\u002Fsim-to-real",{"title":5,"description":90},"glossary\u002Fzh\u002Fsim-to-real",null,"SIqW14dLpX-sh74cSNPCcadVtbF1g-rBDc9vkNN_Wkc",1783425867020]