[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-80912":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":9,"language":10,"languages":9,"totalLinesOfCode":9,"stars":11,"forks":12,"watchers":13,"openIssues":14,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":15,"stars7d":16,"stars30d":17,"stars90d":15,"forks30d":15,"starsTrendScore":15,"compositeScore":18,"rankGlobal":9,"rankLanguage":9,"license":19,"archived":20,"fork":20,"defaultBranch":21,"hasWiki":22,"hasPages":20,"topics":23,"createdAt":9,"pushedAt":9,"updatedAt":24,"readmeContent":25,"aiSummary":26,"trendingCount":15,"starSnapshotCount":15,"syncStatus":27,"lastSyncTime":28,"discoverSource":29},80912,"Go2_ARX_mjlab","Czy213hd\u002FGo2_ARX_mjlab","Czy213hd","MJLAB reinforcement learning for Go2 quadruple robot with ARX-L5 arm.",null,"Python",39,8,33,1,0,5,6,2.86,"Apache License 2.0",false,"main",true,[],"2026-06-12 02:04:08","# Go2_ARX_mjlab\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"docs\u002Fassets\u002Fpic1.png\" alt=\"Go2_ARX_mjlab\" width=\"70%\" \u002F>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"docs\u002Fassets\u002Fpic2.png\" width=\"24%\" \u002F>\n  \u003Cimg src=\"docs\u002Fassets\u002Fpic3.png\" width=\"24%\" \u002F>\n  \u003Cimg src=\"docs\u002Fassets\u002Fpic4.png\" width=\"24%\" \u002F>\n\u003C\u002Fp>\n\nGo2_ARX_mjlab adds a Unitree Go2 + ARX L5 arm robot to\n[mjlab](https:\u002F\u002Fgithub.com\u002Fmujocolab\u002Fmjlab), with reinforcement-learning tasks\nfor base velocity tracking and end-effector pose tracking.\n\nThis repository includes:\n\n- a combined Go2 + ARX L5 MuJoCo XML model\n- position-control actions for the Go2 legs and ARX arm\n- end-effector pose commands for the arm\n- base velocity commands for the quadruped\n- flat and rough terrain task registrations\n\n## Tasks\n\nThe added task IDs are:\n\n```bash\nMjlab-Velocity-Flat-Go2arm\nMjlab-Velocity-Rough-Go2arm\n```\n\nMain implementation files:\n\n```text\nsrc\u002Fmjlab\u002Fasset_zoo\u002Frobots\u002Fgo2arm\u002F\nsrc\u002Fmjlab\u002Ftasks\u002Fvelocity\u002Fconfig\u002Fgo2arm\u002F\nsrc\u002Fmjlab\u002Ftasks\u002Fvelocity\u002Fmdp\u002Fgo2arm_lab.py\n```\n\n## Installation\n\nThis project uses the same setup as mjlab. From the repository root:\n\n```bash\nuv sync\n```\n\nTraining requires an NVIDIA GPU.\n\n## Sanity Check\n\nRun the environment with zero actions:\n\n```bash\nuv run play Mjlab-Velocity-Flat-Go2arm \\\n  --agent zero \\\n  --viewer viser \\\n  --num-envs 1\n```\n\nRun the environment with random actions:\n\n```bash\nuv run play Mjlab-Velocity-Flat-Go2arm \\\n  --agent random \\\n  --viewer viser \\\n  --num-envs 1\n```\n\n## Training\n\nYou can use either Weights & Biases or TensorBoard for logging. The commands\nbelow use TensorBoard for demonstration.\n\nTrain on flat terrain:\n\n```bash\nuv run train Mjlab-Velocity-Flat-Go2arm \\\n  --env.scene.num-envs 4096 \\\n  --agent.logger tensorboard\n```\n\nTrain on rough terrain:\n\n```bash\nuv run train Mjlab-Velocity-Rough-Go2arm \\\n  --env.scene.num-envs 4096 \\\n  --agent.logger tensorboard\n```\n\nTo resume from a local checkpoint:\n\n```bash\nuv run train Mjlab-Velocity-Flat-Go2arm \\\n  --env.scene.num-envs 4096 \\\n  --agent.resume True \\\n  --agent.load-run RUN_DIRECTORY_NAME \\\n  --agent.load-checkpoint model_1000.pt \\\n  --agent.logger tensorboard\n```\n\n## Play a Trained Policy\n\nPlay a checkpoint:\n\n```bash\nuv run play Mjlab-Velocity-Flat-Go2arm \\\n  --checkpoint-file \u002Fpath\u002Fto\u002Fmodel.pt \\\n  --viewer viser \\\n  --num-envs 1\n```\n\n\nFor visualization-only debugging, you can disable terminations:\n\n```bash\nuv run play Mjlab-Velocity-Flat-Go2arm \\\n  --checkpoint-file \u002Fpath\u002Fto\u002Fmodel.pt \\\n  --viewer viser \\\n  --num-envs 1 \\\n  --no-terminations True\n```\n\n## Deployment\n\nDeployment support is currently in progress. The repository includes native\nMuJoCo sim-to-sim scripts for checkpoint sanity checks before real-robot\ndeployment.\n\nRun a checkpoint in native MuJoCo with fixed commands:\n\n```bash\nuv run python deploy\u002Fsimulation\u002Fsim2sim.py \\\n  --checkpoint \u002Fpath\u002Fto\u002Fmodel.pt \\\n  --lin-vel-x 0.2 \\\n  --lin-vel-y 0.0 \\\n  --ang-vel-z 0.0 \\\n  --ee-x 0.48 \\\n  --ee-y 0.0 \\\n  --ee-z 0.36\n```\n\nRun the keyboard-controlled sim-to-sim viewer:\n\n```bash\nuv run python deploy\u002Fsimulation\u002Fsim2sim_keyboard.py \\\n  --checkpoint \u002Fpath\u002Fto\u002Fmodel.pt\n```\n\nKeyboard commands are read from the terminal while the MuJoCo window is used for\nvisualization:\n\n```text\nW \u002F S: base forward velocity +\u002F-\nA \u002F D: base lateral velocity +\u002F-\nQ \u002F E: base yaw velocity +\u002F-\nI \u002F K: EE x +\u002F-\nJ \u002F L: EE y +\u002F-\nU \u002F O: EE z +\u002F-\nR \u002F F: EE pitch offset +\u002F-\nT \u002F G: EE yaw offset +\u002F-\nZ \u002F X: EE roll +\u002F-\nSpace: zero base velocity\nC: reset EE command\nV: print current command\nEsc or Ctrl-C: quit\n```\n\nThe sim-to-sim scripts visualize the target end-effector pose and the measured\nend-effector pose in the MuJoCo viewer.\n\n## Commands and Actions\n\nThe policy receives two commands:\n\n- `base_velocity`: desired base linear and yaw velocity\n- `ee_pose`: desired end-effector pose, represented as\n  `(x, y, z, qw, qx, qy, qz)`\n\nThe action space has 18 dimensions:\n\n- 12 Go2 leg joint position targets\n- 6 ARX L5 arm joint position targets\n\n## Assets\n\nRobot assets include Unitree Go2 and ARX L5 resources. See the included license\nfiles:\n\n```text\nsrc\u002Fmjlab\u002Fasset_zoo\u002Frobots\u002Fgo2arm\u002Fxmls\u002Funitree_go2\u002FLICENSE\nsrc\u002Fmjlab\u002Fasset_zoo\u002Frobots\u002Fgo2arm\u002Fxmls\u002Farx_l5\u002FLICENSE\n```\n\n\n## Contact\u002FAuthor\n\nIf you have questions, issues, or suggestions, please open a GitHub issue or\ncontact the maintainer:\n\n- Zhongyu Chen\n- Email: chenzhongyu56@gmail.com\n\n## Acknowledgments\n\nThis project is built on top of the\n[mjlab](https:\u002F\u002Fgithub.com\u002Fmujocolab\u002Fmjlab) framework. Many thanks to the mjlab\nauthors and contributors for making this work possible.\n\nThe reinforcement-learning task design and algorithm setup reference\n[Go2Arm_Lab](https:\u002F\u002Fgithub.com\u002FzzzJie-Robot\u002FGo2Arm_Lab). We sincerely thank\nthe authors for their open-source work.\n\n## License\n\nThis repository is based on mjlab and keeps the original Apache-2.0 license.\nSee [LICENSE](LICENSE).\n\nThird-party assets and code retain their original licenses. In particular,\ncheck the license files bundled with the Go2 and ARX L5 assets before using them\nin commercial or redistributed projects.\n\nIf you use the underlying mjlab framework in research, please also cite the\noriginal mjlab project.\n","Go2_ARX_mjlab 项目旨在为 Unitree Go2 四足机器人搭配 ARX-L5 机械臂提供强化学习任务支持。该项目通过结合MuJoCo物理引擎，实现了对机器人腿部和机械臂的位置控制、末端执行器姿态追踪以及四足机器人的基础速度控制等功能。其核心功能包括在平坦和复杂地形上进行速度跟踪任务，并提供了详细的训练和测试脚本。适合于需要研究或开发基于强化学习的移动操作机器人系统的场景，特别是在探索如何提高机器人在不同环境下的适应性和操作精度方面。此外，该项目还支持使用TensorBoard等工具进行训练过程的可视化监控。",2,"2026-06-11 04:02:47","CREATED_QUERY"]