LAFAN1_Retargeting_Dataset

# LAFAN1 Retargeting Dataset

<video controls autoplay src="https://cdn-uploads.huggingface.co/production/uploads/67639932ad38702e6c8d16d9/V7InyG4CAh5NhUXILTK9b.mp4"></video>

To make the motion of humanoid robots more natural, we retargeted [LAFAN1](https://github.com/ubisoft/ubisoft-laforge-animation-dataset) motion capture data to [Unitree](https://www.unitree.com/)'s humanoid robots, supporting three models: [H1, H1_2](https://www.unitree.com/h1), and [G1](https://www.unitree.com/g1). This retargeting was achieved through numerical optimization based on [Interaction Mesh](https://ieeexplore.ieee.org/document/6651585) and IK, considering end-effector pose constraints, as well as joint position and velocity constraints, to prevent foot slippage. It is important to note that the retargeting only accounted for kinematic constraints and did not include dynamic constraints or actuator limitations. As a result, the robot cannot perfectly execute the retargeted trajectories.

# How to visualize robot trajectories?

```shell
# Step 1: Set up a Conda virtual environment
conda create -n retarget python=3.10
conda activate retarget

# Step 2: Install dependencies
conda install pinocchio -c conda-forge
pip install numpy rerun-sdk==0.22.0 trimesh

# Step 3: Run the script
python rerun_visualize.py
# run the script with parameters:
# python rerun_visualize.py --file_name dance1_subject2 --robot_type [g1|h1|h1_2]
```

# Dataset Collection Pipeline


This database stores the retargeted trajectories in CSV format. Each row in the CSV file corresponds to the original motion capture data for each frame, recording the configurations of all joints in the humanoid robot in the following order:

```txt
The Order of Configuration
G1: (30 FPS)
root_joint(XYZQXQYQZQW)
left_hip_pitch_joint
left_hip_roll_joint
left_hip_yaw_joint
left_knee_joint
left_ankle_pitch_joint
left_ankle_roll_joint
right_hip_pitch_joint
right_hip_roll_joint
right_hip_yaw_joint
right_knee_joint
right_ankle_pitch_joint
right_ankle_roll_joint
waist_yaw_joint
waist_roll_joint
waist_pitch_joint
left_shoulder_pitch_joint
left_shoulder_roll_joint
left_shoulder_yaw_joint
left_elbow_joint
left_wrist_roll_joint
left_wrist_pitch_joint
left_wrist_yaw_joint
right_shoulder_pitch_joint
right_shoulder_roll_joint
right_shoulder_yaw_joint
right_elbow_joint
right_wrist_roll_joint
right_wrist_pitch_joint
right_wrist_yaw_joint
H1_2: (30 FPS)
root_joint(XYZQXQYQZQW)
left_hip_yaw_joint
left_hip_pitch_joint
left_hip_roll_joint
left_knee_joint
left_ankle_pitch_joint
left_ankle_roll_joint
right_hip_yaw_joint
right_hip_pitch_joint
right_hip_roll_joint
right_knee_joint
right_ankle_pitch_joint
right_ankle_roll_joint
torso_joint
left_shoulder_pitch_joint
left_shoulder_roll_joint
left_shoulder_yaw_joint
left_elbow_joint
left_wrist_roll_joint
left_wrist_pitch_joint
left_wrist_yaw_joint
right_shoulder_pitch_joint
right_shoulder_roll_joint
right_shoulder_yaw_joint
right_elbow_joint
right_wrist_roll_joint
right_wrist_pitch_joint
right_wrist_yaw_joint
H1: (30 FPS)
root_joint(XYZQXQYQZQW)
left_hip_yaw_joint
left_hip_roll_joint
left_hip_pitch_joint
left_knee_joint
left_ankle_joint
right_hip_yaw_joint
right_hip_roll_joint
right_hip_pitch_joint
right_knee_joint
right_ankle_joint
torso_joint
left_shoulder_pitch_joint
left_shoulder_roll_joint
left_shoulder_yaw_joint
left_elbow_joint
right_shoulder_pitch_joint
right_shoulder_roll_joint
right_shoulder_yaw_joint
right_elbow_joint
```

[LAFAN1](https://github.com/ubisoft/ubisoft-laforge-animation-dataset) is licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International Public License (unlike the code, which is licensed under MIT).

# LAFAN1 运动重定向数据集

<video controls autoplay src="https://cdn-uploads.huggingface.co/production/uploads/67639932ad38702e6c8d16d9/V7InyG4CAh5NhUXILTK9b.mp4"></video>

为使类人机器人的运动更趋自然，我们将[LAFAN1](https://github.com/ubisoft/ubisoft-laforge-animation-dataset)运动捕捉数据重定向至[宇树（Unitree）](https://www.unitree.com/)旗下的三款类人机器人，分别为[H1、H1_2](https://www.unitree.com/h1)与[G1](https://www.unitree.com/g1)。本次重定向采用基于[Interaction Mesh](https://ieeexplore.ieee.org/document/6651585)与逆运动学（IK）的数值优化方法，同时考虑末端执行器位姿约束、关节位置与速度约束，以避免足部滑动。需要特别说明的是，本次重定向仅考虑了运动学约束，未纳入动力学约束与执行器限制条件，因此机器人无法完美复现重定向后的运动轨迹。

# 如何可视化机器人运动轨迹？

shell
# 步骤1：配置Conda虚拟环境
conda create -n retarget python=3.10
conda activate retarget

# 步骤2：安装依赖包
conda install pinocchio -c conda-forge
pip install numpy rerun-sdk==0.22.0 trimesh

# 步骤3：运行脚本
python rerun_visualize.py
# 带参数运行脚本：
# python rerun_visualize.py --file_name dance1_subject2 --robot_type [g1|h1|h1_2]

# 数据集采集流程

![image/png]("https://cdn-uploads.huggingface.co/production/uploads/67639932ad38702e6c8d16d9/_tAr3zwPotJGaUqxe2u4p.png")

本数据集以CSV格式存储重定向后的运动轨迹。CSV文件中的每一行对应一帧原始运动捕捉数据，按照以下顺序记录类人机器人所有关节的构型：

txt
关节构型顺序
G1：帧率30 FPS
根关节（XYZQXQYQZQW）
左髋俯仰关节
左髋侧滚关节
左髋偏航关节
左膝关节
左踝俯仰关节
左踝侧滚关节
右髋俯仰关节
右髋侧滚关节
右髋偏航关节
右膝关节
右踝俯仰关节
右踝侧滚关节
腰偏航关节
腰侧滚关节
腰俯仰关节
左肩俯仰关节
左肩侧滚关节
左肩偏航关节
左肘关节
左腕侧滚关节
左腕俯仰关节
左腕偏航关节
右肩俯仰关节
右肩侧滚关节
右肩偏航关节
右肘关节
右腕侧滚关节
右腕俯仰关节
右腕偏航关节
H1_2：帧率30 FPS
根关节（XYZQXQYQZQW）
左髋偏航关节
左髋俯仰关节
左髋侧滚关节
左膝关节
左踝俯仰关节
左踝侧滚关节
右髋偏航关节
右髋俯仰关节
右髋侧滚关节
右膝关节
右踝俯仰关节
右踝侧滚关节
躯干关节
左肩俯仰关节
左肩侧滚关节
左肩偏航关节
左肘关节
左腕侧滚关节
左腕俯仰关节
左腕偏航关节
右肩俯仰关节
右肩侧滚关节
右肩偏航关节
右肘关节
右腕侧滚关节
右腕俯仰关节
右腕偏航关节
H1：帧率30 FPS
根关节（XYZQXQYQZQW）
左髋偏航关节
左髋侧滚关节
左髋俯仰关节
左膝关节
左踝关节
右髋偏航关节
右髋侧滚关节
右髋俯仰关节
右膝关节
右踝关节
躯干关节
左肩俯仰关节
左肩侧滚关节
左肩偏航关节
左肘关节
右肩俯仰关节
右肩侧滚关节
右肩偏航关节
右肘关节

[LAFAN1](https://github.com/ubisoft/ubisoft-laforge-animation-dataset)采用知识共享署名-非商业性使用-禁止演绎4.0国际公共许可协议进行授权（与采用MIT许可协议的代码不同）。

AI搜集汇总

背景与挑战

背景概述

该数据集是将LAFAN1运动捕捉数据重定向到Unitree人形机器人（支持H1、H1_2和G1模型）的轨迹数据，旨在提升机器人运动的自然度。重定向过程通过数值优化考虑运动学约束，但未包含动态限制，因此机器人无法完美执行轨迹。数据以CSV格式存储，每帧记录机器人关节配置，适用于机器人运动研究和可视化应用。