[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-83113":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":16,"subscribersCount":16,"size":16,"stars1d":17,"stars7d":18,"stars30d":18,"stars90d":16,"forks30d":16,"starsTrendScore":19,"compositeScore":20,"rankGlobal":10,"rankLanguage":10,"license":10,"archived":21,"fork":21,"defaultBranch":22,"hasWiki":23,"hasPages":21,"topics":24,"createdAt":10,"pushedAt":25,"updatedAt":26,"readmeContent":27,"aiSummary":28,"trendingCount":16,"starSnapshotCount":16,"syncStatus":14,"lastSyncTime":29,"discoverSource":30},83113,"Humanoid-GPT","GalaxyGeneralRobotics\u002FHumanoid-GPT","GalaxyGeneralRobotics","[CVPR 2026] Humanoid-GPT: Scaling Data and Structure for Zero-Shot Motion Tracking","https:\u002F\u002Fqizekun.github.io\u002FHumanoid-GPT\u002F",null,"Python",213,11,2,1,0,9,93,70,87.42,false,"main",true,[],"2026-06-08 22:23:57","2026-06-10 15:02:18","\u003Cdiv align=\"center\">\n\n# 🤖 Humanoid-GPT\n\n### [CVPR 2026] Humanoid Generative Pre-Training for Zero-Shot Motion Tracking\n\n\u003Cp align=\"center\">\n  \u003Ca href=\"https:\u002F\u002Fcvpr.thecvf.com\u002FConferences\u002F2026\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FCVPR-2026-4b44ce.svg?style=flat-square\" alt=\"CVPR 2026\">\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.03985\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2606.03985-b31b1b.svg?style=flat-square\" alt=\"arXiv\">\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Fqizekun.github.io\u002FHumanoid-GPT\u002F\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProject-Page-blue.svg?style=flat-square\" alt=\"Project Page\">\u003C\u002Fa>\n  \u003Ca href=\"LICENSE\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-Apache%202.0-green.svg?style=flat-square\" alt=\"License\">\u003C\u002Fa>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"storage\u002Fassets\u002Fteaser.png\" width=\"100%\" alt=\"Humanoid-GPT Teaser\">\n\u003C\u002Fp>\n\n\u003C\u002Fdiv>\n\n---\n\n## 📖 Overview\n\n**Humanoid-GPT** is the first **GPT-style humanoid motion Transformer** trained with causal attention on a billion-scale motion corpus for whole-body control. Unlike prior shallow MLP trackers constrained by scarce data and an agility–generalization trade-off, Humanoid-GPT is pre-trained on a **2B-frame retargeted corpus** that unifies all major mocap datasets with large-scale in-house recordings.\n\n\u003Cdetails>\n\u003Csummary>\u003Cb>🔬 Key Contributions\u003C\u002Fb>\u003C\u002Fsummary>\n\n- **Billion-Scale Pre-Training**: First to scale humanoid motion learning to 2B frames\n- **GPT-Style Architecture**: Causal Transformer with Rotary Position Embeddings (RoPE)\n- **Zero-Shot Generalization**: Track arbitrary unseen motions without fine-tuning\n\n\u003C\u002Fdetails>\n\n### ✨ Highlights\n\n| Feature             | Description                                                               |\n|---------------------|---------------------------------------------------------------------------|\n| 🧠 **Architecture** | Causal Transformer with RoPE, supporting variable-length motion sequences |\n| 📊 **Scale**        | Pre-trained on 2B motion frames from unified mocap datasets               |\n| 🎯 **Zero-Shot**    | Unprecedented generalization to unseen motions and tasks                  |\n| 🤖 **Platform**     | Optimized for Unitree G1 humanoid robot (29 DOF whole-body)               |\n| ⚡  **Speed**        | GPU-accelerated simulation with MuJoCo-MJX                                |\n\n---\n\n## 📦 Installation\n\n### Prerequisites\n\n- NVIDIA GPU with CUDA 12.x\n- **MacOS** is also supported for testing if you skip **jax[cuda12]** and use **mjpython** (e.g. `mjpython -m scripts.app`).\n- Conda \u002F Miniconda\n\n### Quick Start\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002Fqizekun\u002FHumanoid-GPT.git\ncd Humanoid-GPT\n\nconda create -n h-gpt python=3.12 -y\nconda activate h-gpt\n\npip install -e \".[cuda]\"     # or \".[cpu]\" on MacOS, or \".\" for real robot deploy-only\n```\n\nOn MacOS, use `mjpython` instead of `python` for the MuJoCo viewer (e.g. `mjpython -m scripts.app`).\n\n### 🔧 G1 Hardware Version\n\nWe support multiple Unitree G1 hardware versions via the `G1_VERSION` env var (default `5010`). The asset folder `storage\u002Fassets\u002Funitree_g1_${G1_VERSION}\u002F` is selected automatically:\n\n```bash\nG1_VERSION=5010 python -m scripts.inference ...                   # default: 5010\n```\n\n---\n\n## 🚀 Inference & Evaluation\n\nA pre-trained tracking policy (`.onnx`) and a sample trajectory under\n`storage\u002Ftest\u002F` are all you need to get started.\n\n```bash\n# Interactive Gradio demo\npython -m scripts.app\n\n# Track a single motion \u002F a folder of motions\npython -m scripts.inference --load_path storage\u002Fckpts\u002Fpns_wo_priv216.onnx --mocap_path storage\u002Ftest\n\n# Parallel evaluation over a folder of trajectories\npython -m scripts.eval_parallel --load_path storage\u002Fckpts\u002Fpns_wo_priv216.onnx \\\n    --mocap_path storage\u002Ftest --workers 32 --privileged\n\n# Visualize a reference trajectory\npython -m scripts.vis --mocap_path storage\u002Ftest\n```\n\nThe expected motion format is a `.npz` containing either `qpos` directly, or\n`root_pos` \u002F `root_rot` \u002F `dof_pos` arrays. To convert retargeted mocap into\nthe keypoint representation the policy consumes:\n\n```bash\npython tracking\u002Fconvert_qpos2kpt.py --mocap_npz \u003Cmocap_path.npz> --debug   # single file (debug viz)\npython tracking\u002Fconvert_parallel.py --src_dir \u003Cin_dir> --save_dir \u003Cout_dir> --num_workers 32\n```\n\n---\n\n## 🤖 Real-Robot Deployment\n\nDeployment on Unitree G1 is split into sub-modules under `deploy\u002F` — start with\n**[`deploy\u002FDEPLOY.md`](deploy\u002FDEPLOY.md)** for install \u002F SDK setup, then:\n\n```bash\n# Simulation\npython -m deploy.play_track --track-dir storage\u002Ftest\n\n# Real robot\npython -m deploy.play_track --real --net \u003Cnic_name>\n```\n\n- 🖥️ [`onboard_deploy\u002F`](deploy\u002Fonboard_deploy\u002FDEPLOY_ONBOARD.md) — on-board (Jetson Orin) deploy.\n- 🖥️ `onboard_deploy_wo_GMR\u002F` — on-board variant that streams retargeting from a host.\n- ✋ [`brainco\u002F`](deploy\u002Fbrainco\u002FBRAINCO.md) — BrainCo dexterous-hand tracking variant.\n\n---\n\n## 📁 Project Structure\n\n```\nHumanoid-GPT\u002F\n├── 📂 tracking\u002F   # Inference core: constants, infer_utils, ONNX policy wrapper (policy.py),\n│                  # keypoint conversion (convert_qpos2kpt.py) and tracking metrics\n├── 📂 scripts\u002F    # inference.py · eval_parallel.py · vis.py · app.py (gradio demo)\n├── 📂 deploy\u002F     # Real-robot deployment — see deploy\u002FDEPLOY.md\n│   ├── onboard_deploy\u002F        # On-board (Jetson) SSH deployment\n│   ├── onboard_deploy_wo_GMR\u002F # On-board variant with host-side retargeting\n│   └── brainco\u002F               # BrainCo dexterous-hand tracking variant\n├── 📂 projects\u002F   # Optional side modules\n│   ├── hme\u002F                  # Harmonic Motion Encoder (Periodic Autoencoder)\n│   ├── gqs\u002F                  # General Quality Selection (physics + diversity scoring)\n│   └── tracking_transformer\u002F # Transformer tracking policy (inference \u002F deploy)\n├── 📂 utils\u002F      # MuJoCo \u002F MJX simulation, transforms, video rendering\n└── 📂 storage\u002F    # Assets, configs, sample trajectory, released checkpoints\n```\n\n---\n\n## 📚 Citation\n\n```bibtex\n@article{humanoid-gpt26,\n    title     = {Humanoid-GPT: Humanoid Generative Pre-Training for Zero-Shot Motion Tracking},\n    author    = {Qi, Zekun and Chen, Xuchuan and others},\n    journal   = {arXiv preprint arXiv:2606.03985},\n    year      = {2026}\n}\n```\n\n---\n\n## 📄 License · Acknowledgments\n\nLicensed under **Apache 2.0**. Built on top of [MuJoCo](https:\u002F\u002Fmujoco.org\u002F), [Brax](https:\u002F\u002Fgithub.com\u002Fgoogle\u002Fbrax) and the [Unitree](https:\u002F\u002Fwww.unitree.com\u002F) G1 platform.\n","Humanoid-GPT 是一个基于GPT风格的人形运动Transformer，通过因果注意力机制在数十亿帧的运动数据集上进行预训练，以实现全身控制。该项目采用因果Transformer架构并结合旋转位置嵌入（RoPE），支持变长运动序列处理，并在20亿帧统一动作捕捉数据集上进行了预训练，实现了对未见过动作的零样本泛化能力。特别适合需要高精度、无需额外微调即可适应新任务或动作的场景，如机器人技术中的动态运动跟踪与控制。此外，该模型针对Unitree G1人形机器人进行了优化，支持29自由度的全身体控制，并且可以通过GPU加速提高仿真速度。","2026-06-11 04:10:09","CREATED_QUERY"]