ami-iit/amp-dataset

Retargeted Robot Motion Dataset

This dataset provides retargeted motion capture sequences for a variety of robotic platforms. The motion data is derived from the CMU Motion Capture Database and includes a wide range of motion types beyond locomotion — such as gestures, interactions, and full-body activities.

The data has been adapted to match the kinematic structure of specific robots, enabling its use in tasks such as:

• Imitation learning
• Reinforcement learning
• Motion analysis
• Skill learning from demonstration

---

📁 Dataset Format

Each .npy file contains a dictionary with the following keys:

• joints_list: list of joint names
• joint_positions: per-frame joint angle vectors
• root_position: per-frame 3D position of the robot’s base
• root_quaternion: per-frame orientation as quaternions (xyzw)
• fps: sampling rate of the motion (frames per second)
• duration: total sequence duration in seconds
• r_forearm, l_forearm: optional 3D positions of the forearms
• r_sole, l_sole: optional 3D positions of the soles of the feet

All lists should have consistent lengths across time frames.

Example

{
  "joints_list": ["joint_1", "joint_2", ...],
  "joint_positions": [np.array([...]), ...],
  "root_position": [np.array([x, y, z]), ...],
  "root_quaternion": [np.array([x, y, z, w]), ...],
  "fps": 120.0,
  "duration": 3.2
}

---

📂 Repository Structure

The dataset repository is organized by robot model:

.
├── ergocub/
│   ├── wave_hand.npy
│   ├── walk_forward.npy
│   └── ...
├── another_robot/
│   └── ...
├── README.md

Each folder contains .npy files with motion data retargeted to that specific robot's model.

---

📄 License

Original motion capture data © Carnegie Mellon University.
Retargeted datasets are provided for research purposes under the BSD 3-Clause license.