Cryyzz/Lafan1_Retarget_To_BUMI-GMR

---
license: cc-by-nc-4.0
task_categories:
- robotics
tags:
- motion-retargeting
- humanoid-robot
- bvh
- motion-capture
- mujoco
- lafan1
- bumi
language:
- en
pretty_name: LAFAN1 to BUMI GMR (Generalized Motion Retargeting)
size_categories:
- 100M<n<1B
---

# LAFAN1 to BUMI — Generalized Motion Retargeting Dataset

## Dataset Description

This dataset contains retargeted motion data transferred from the **LAFAN1** human motion capture dataset onto the **BUMI V3.0** bipedal humanoid robot using a two-pass inverse kinematics (IK) pipeline. The resulting joint trajectories are ready to use for robot learning, imitation learning, and motion control research.

- **Repository:** `Cryyzz/lafan1-to-bumi-gmr`
- **License:** CC-BY-NC-4.0
- **Robot:** BUMI V3.0 — 21-DOF bipedal humanoid
- **Source motion:** LAFAN1 (BVH format)

---

## Files

| File | Size | Description |
|---|---|---|
| `lafan1-to-bumi-gmr.zip` | 102 MB | Retargeted joint trajectories for all LAFAN1 sequences |
| `bvh_lafan1_to_bumi.json` | 4.61 KB | Retargeting configuration: IK mapping tables and scale factors |

### Robot Model Files

The following model files are used for simulation and retargeting. Please refer to the BUMI robot platform for the full model package (meshes + URDF/MJCF).

| File | Format | Description |
|---|---|---|
| `BUMI_V3_0_collision_v4.urdf` | URDF | Robot description for ROS / other toolchains |
| `bumi_v3_v4.xml` | MuJoCo MJCF | Robot description for MuJoCo physics simulation |

---

## Robot: BUMI V3.0

BUMI V3.0 is a 21-DOF bipedal humanoid robot with symmetric left/right limb design.

**Kinematic structure:**

```
base_link
├── waist_yaw_link
│ ├── l_arm_pitch → l_arm_roll → l_arm_yaw → l_elbow_pitch → [l_hand_link]
│ └── r_arm_pitch → r_arm_roll → r_arm_yaw → r_elbow_pitch → [r_hand_link]
├── l_leg_pitch → l_leg_roll → l_leg_yaw → l_knee_pitch → l_ankle_pitch → l_ankle_roll
└── r_leg_pitch → r_leg_roll → r_leg_yaw → r_knee_pitch → r_ankle_pitch → r_ankle_roll
```

**Joint count by body part:**

| Body Part | Joints | DOF |
|---|---|---|
| Waist | waist_yaw | 1 |
| Left Arm | l_arm_pitch, l_arm_roll, l_arm_yaw, l_elbow_pitch | 4 |
| Right Arm | r_arm_pitch, r_arm_roll, r_arm_yaw, r_elbow_pitch | 4 |
| Left Leg | l_leg_pitch, l_leg_roll, l_leg_yaw, l_knee_pitch, l_ankle_pitch, l_ankle_roll | 6 |
| Right Leg | r_leg_pitch, r_leg_roll, r_leg_yaw, r_knee_pitch, r_ankle_pitch, r_ankle_roll | 6 |
| **Total** | | **21** |

---

## Retargeting Pipeline

The retargeting uses a **two-pass IK** strategy defined in `bvh_lafan1_to_bumi.json`.

### Scale Factors (`human_scale_table`)

Human bone lengths are scaled down to match BUMI's proportions before IK solving (assuming a reference human height of 1.8 m):

| Body Segment | Scale |
|---|---|
| Hips, Spine, Legs | 0.55 |
| Arms, Forearms, Hands | 0.60 |

### IK Pass 1 (`ik_match_table1`) — Coarse Alignment

Emphasises **rotation matching** across the whole body. The torso (`waist_yaw_link`) is given a high rotation weight (100) to anchor the upper body orientation first. Arm and leg segments are aligned with moderate weights.

### IK Pass 2 (`ik_match_table2`) — Fine End-Effector Matching

Increases **position weights** on the root (`base_link`) and end-effectors (`l_hand_link`, `r_hand_link`, `l_ankle_roll_link`, `r_ankle_roll_link`) to achieve accurate foot placement and hand positioning.

### IK Entry Format

```
"robot_link": ["human_bone", position_weight, rotation_weight, [pos_offset_x, y, z], [quat_w, x, y, z]]
```

---

## Source Dataset

**LAFAN1** — A large-scale human motion capture dataset for locomotion and action research.

Human skeleton bones used in this retargeting:
`Hips`, `Spine2`, `LeftUpLeg`, `RightUpLeg`, `LeftLeg`, `RightLeg`, `LeftFootMod`, `RightFootMod`, `LeftArm`, `RightArm`, `LeftForeArm`, `RightForeArm`, `LeftHand`, `RightHand`

---

## Intended Uses

- Humanoid robot imitation learning
- Motion control policy training
- Sim-to-real transfer research
- Benchmarking motion retargeting methods

## Out-of-Scope Uses

- Commercial use (see CC-BY-NC-4.0 license)
- Direct deployment on physical robots without safety validation

---

## Citation

If you use this dataset in your research, please cite the LAFAN1 dataset and the BUMI robot platform:

```bibtex
@dataset{cryyzz2025lafan1bumi,
author = {Cryyzz},
title = {LAFAN1 to BUMI Generalized Motion Retargeting Dataset},
year = {2025},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/Cryyzz/lafan1-to-bumi-gmr}
}
```

---

---

# LAFAN1 到 BUMI — 通用动作重定向数据集

## 数据集简介

本数据集包含从 **LAFAN1** 人体动作捕捉数据集重定向到 **BUMI V3.0** 双足人形机器人的动作数据，采用两轮逆运动学（IK）流程生成。输出的关节轨迹可直接用于机器人学习、模仿学习和运动控制研究。

- **仓库：** `Cryyzz/lafan1-to-bumi-gmr`
- **许可证：** CC-BY-NC-4.0
- **机器人：** BUMI V3.0 — 21 自由度双足人形机器人
- **源动作数据：** LAFAN1（BVH 格式）

---

## 文件说明

| 文件 | 大小 | 说明 |
|---|---|---|
| `lafan1-to-bumi-gmr.zip` | 102 MB | 所有 LAFAN1 序列的重定向关节轨迹 |
| `bvh_lafan1_to_bumi.json` | 4.61 KB | 重定向配置：IK 映射表与缩放比例 |

### 机器人模型文件

以下模型文件用于仿真和重定向计算，完整模型包（含 STL 网格 + URDF/MJCF）请参阅 BUMI 机器人平台：

| 文件 | 格式 | 说明 |
|---|---|---|
| `BUMI_V3_0_collision_v4.urdf` | URDF | 用于 ROS 及其他工具链的机器人描述文件 |
| `bumi_v3_v4.xml` | MuJoCo MJCF | 用于 MuJoCo 物理仿真的机器人描述文件 |

---

## 机器人：BUMI V3.0

BUMI V3.0 是一款具有 21 个自由度的双足人形机器人，采用左右对称肢体设计。

**运动学链结构：**

```
base_link
├── waist_yaw_link
│ ├── l_arm_pitch → l_arm_roll → l_arm_yaw → l_elbow_pitch → [l_hand_link]
│ └── r_arm_pitch → r_arm_roll → r_arm_yaw → r_elbow_pitch → [r_hand_link]
├── l_leg_pitch → l_leg_roll → l_leg_yaw → l_knee_pitch → l_ankle_pitch → l_ankle_roll
└── r_leg_pitch → r_leg_roll → r_leg_yaw → r_knee_pitch → r_ankle_pitch → r_ankle_roll
```

**各部位自由度：**

| 部位 | 关节 | 自由度 |
|---|---|---|
| 腰部 | waist_yaw | 1 |
| 左臂 | l_arm_pitch, l_arm_roll, l_arm_yaw, l_elbow_pitch | 4 |
| 右臂 | r_arm_pitch, r_arm_roll, r_arm_yaw, r_elbow_pitch | 4 |
| 左腿 | l_leg_pitch, l_leg_roll, l_leg_yaw, l_knee_pitch, l_ankle_pitch, l_ankle_roll | 6 |
| 右腿 | r_leg_pitch, r_leg_roll, r_leg_yaw, r_knee_pitch, r_ankle_pitch, r_ankle_roll | 6 |
| **合计** | | **21** |

---

## 重定向流程

重定向采用 `bvh_lafan1_to_bumi.json` 中定义的**两轮 IK** 策略。

### 缩放比例（`human_scale_table`）

在 IK 求解前，将人体骨骼长度按比例缩放至 BUMI 的体型（参考人体身高 1.8 m）：

| 身体部位 | 缩放比例 |
|---|---|
| 髋部、脊柱、腿部 | 0.55 |
| 手臂、前臂、手部 | 0.60 |

### 第一轮 IK（`ik_match_table1`）— 粗粒度对齐

强调全身**旋转匹配**。躯干（`waist_yaw_link`）旋转权重设为 100，优先锁定上身姿态方向；手臂和腿部以中等权重进行对齐。

### 第二轮 IK（`ik_match_table2`）— 末端精细匹配

提高根节点（`base_link`）和末端执行器（`l_hand_link`、`r_hand_link`、`l_ankle_roll_link`、`r_ankle_roll_link`）的**位置权重**，实现精确的落脚点和手部位置控制。

### IK 配置格式

```
"机器人连杆": ["人体骨骼", 位置权重, 旋转权重, [位置偏移 x, y, z], [四元数 w, x, y, z]]
```

---

## 源数据集

**LAFAN1** — 大规模人体动作捕捉数据集，覆盖多种运动和动作类别。

本重定向使用的人体骨骼节点：
`Hips`、`Spine2`、`LeftUpLeg`、`RightUpLeg`、`LeftLeg`、`RightLeg`、`LeftFootMod`、`RightFootMod`、`LeftArm`、`RightArm`、`LeftForeArm`、`RightForeArm`、`LeftHand`、`RightHand`

---

## 适用场景

- 人形机器人模仿学习
- 运动控制策略训练
- 仿真到现实（Sim-to-Real）迁移研究
- 动作重定向方法的基准测试

## 不适用场景

- 商业用途（见 CC-BY-NC-4.0 许可证）
- 未经安全验证直接部署到真实机器人上

---

## 引用

如果您在研究中使用了本数据集，请引用 LAFAN1 数据集和 BUMI 机器人平台：

```bibtex
@dataset{cryyzz2025lafan1bumi,
author = {Cryyzz},
title = {LAFAN1 to BUMI Generalized Motion Retargeting Dataset},
year = {2025},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/Cryyzz/lafan1-to-bumi-gmr}
}
```