--- pretty_name: GMR NE01 Paired Motion Dataset license: other tags: - robotics - humanoid-robotics - motion-capture - motion-retargeting - smpl - gmr configs: - config_name: diverse_actions_8h_v3 data_files: - split: train path: manifest.csv --- # GMR NE01 Paired Motion Dataset `diverse_actions_8h_v3` 是一套面向 NE01 24-DoF 人形机器人的成对动作数据。每个样本同时提供重定向后的 GMR robot-motion 和重定向前、已规范化到 GMR 坐标系的 SMPL 数据,并通过根目录的 `manifest.csv` 一一关联。 ## 数据概览 | 项目 | 数值 | | --- | ---: | | 成对样本 | 4376 | | 总时长 | 8.001772 小时 | | Robot-motion | 4376 个 `.pkl`,789.56 MiB | | SMPL | 4376 个 `.npz`,424.09 MiB | | 合计 | 1.19 GiB | | FPS | 50 | | Robot DoF | 24 | 目录结构: ```text . ├── README.md ├── manifest.csv ├── dataset_summary.json ├── robot_motion/ │ └── diverse_actions_8h_v3/ │ └── *.pkl └── smpl/ └── diverse_actions_8h_v3/ └── *.npz ``` 相同 `sample_id` 的 `.pkl` 和 `.npz` 表示同一个动作。`manifest.csv` 同时给出两侧路径、文件大小、SHA-256、是否存在以及 `pair_status`。 ## 数据状态与筛选说明 本版本按 `contact_free` profile 选择,并对 robot-motion 做过关节跳变清理。筛选状态完整保留在 manifest 中: | 策略 | pass | tag | reject | | --- | ---: | ---: | ---: | | Strict | 1048 | 787 | 2541 | | Relaxed dynamic | 1048 | 3221 | 107 | - `primary`: 4269 条。 - `arm_side_fallback`: 107 条;这些样本的 relaxed reject 原因仅为 `arm_side_anomaly`,因类别/时长覆盖需要被限量保留。 - Relaxed dynamic 策略允许部分高动态动作从 strict reject 转为 tag。训练前请根据自己的机器人、控制器和风险容忍度使用 `strict_status`、`relaxed_status`、`selection_tier`、`strict_reject_reasons` 和 `strict_tag_reasons` 再筛选。 - `contact_free` 表示元数据规则未检测到道具、外部支撑或高处跳下依赖,不代表动作已经在真实机器人上通过安全验证。 ## 动作类别 | Category | Samples | Hours | | --- | ---: | ---: | | `upper_body_gesture` | 783 | 1.141761 | | `locomotion_jog_run` | 705 | 1.127533 | | `locomotion_walk` | 573 | 1.342767 | | `jump` | 438 | 0.789611 | | `turn_transition` | 372 | 0.522961 | | `dance` | 339 | 0.725572 | | `idle_stance` | 273 | 0.667583 | | `injury_impaired_gait` | 256 | 0.596956 | | `daily_social_expression` | 225 | 0.338850 | | `obstacle_contact_avoidance` | 223 | 0.396489 | | `exercise_sport` | 91 | 0.148850 | | `other` | 51 | 0.107767 | | `ground_low_posture` | 23 | 0.048611 | | `kick_throw_stoop` | 19 | 0.035417 | | `object_manipulation_carry` | 5 | 0.011044 | ## 下载 安装或临时运行 Hugging Face CLI 后下载完整仓库: ```bash hf download mahanxin/gmr_ne01 \ --repo-type dataset \ --local-dir ./gmr_ne01 ``` 只下载 robot-motion: ```bash hf download mahanxin/gmr_ne01 \ --repo-type dataset \ --include "robot_motion/diverse_actions_8h_v3/**" \ --local-dir ./gmr_ne01 ``` 只下载 SMPL: ```bash hf download mahanxin/gmr_ne01 \ --repo-type dataset \ --include "smpl/diverse_actions_8h_v3/**" \ --local-dir ./gmr_ne01 ``` Python: ```python from huggingface_hub import snapshot_download snapshot_download( repo_id="mahanxin/gmr_ne01", repo_type="dataset", local_dir="./gmr_ne01", ) ``` ## 读取示例 Robot-motion 是 Python pickle 字典。只对可信来源使用 `pickle.load`: ```python import pickle with open( "gmr_ne01/robot_motion/diverse_actions_8h_v3/.pkl", "rb" ) as stream: motion = pickle.load(stream) print(motion.keys()) # fps, root_pos, root_rot, dof_pos, local_body_pos, link_body_list # root_rot 使用 xyzw 四元数顺序;dof_pos.shape == (T, 24) ``` SMPL: ```python import numpy as np path = "gmr_ne01/smpl/diverse_actions_8h_v3/.npz" with np.load(path, allow_pickle=False) as motion: poses = motion["poses"] # (T, 72), axis-angle trans = motion["trans"] # (T, 3) betas = motion["betas"] fps = float(motion["mocap_framerate"]) ``` SMPL 文件是 Sonic Y-up 转换到 GMR Z-up 后的规范化数据,`coord_transform=sonic_yup_to_gmr_zup`,并非原始 Sonic pickle。 ## Manifest 关键字段 - 配对与文件:`sample_id`, `robot_motion_path`, `smpl_path`, `has_robot_motion`, `has_smpl`, `pair_status`, `*_bytes`, `*_sha256` - 动作元数据:`category`, `motion_family`, `actor`, `num_frames`, `fps`, `duration_sec` - 数据情况:`strict_status`, `relaxed_status`, `selection_tier`, `strict_reject_reasons`, `strict_tag_reasons`, `contact_free`, `robot_joint_cleaned` - 格式:`robot_motion_format`, `robot_motion_root_rotation_order`, `smpl_format`, `smpl_gender`, `smpl_coord_transform` 原始本机绝对路径未发布。 ## 可视化与格式文档 可使用 [GMR](https://github.com/YanjieZe/GMR) 的 `scripts/vis_robot_motion.py` 回放 robot-motion。Robot-motion 中的关节顺序与 NE01 XML 一致。 ## License 本仓库暂以 Hugging Face `other` 标记发布。GMR 代码采用 MIT License;数据本身的适用授权与使用范围请由发布者和使用者根据源数据权利另行确认。