README.md

---

task_categories:
- robotics

language:
- en

tags:
- motion-capture
- humanoid-robotics
- human-motion
- physical-ai
- whole-body-control
- nvidia-soma
- unitree-g1
- bvh
- csv
- temporal-annotations
- imitation-learning
- motion-retargeting

pretty_name: Apple Arts Studios Motion Dataset

size_categories:
- 1K<n<10K

license: other
license_name: apple-arts-studios-motion-dataset-license
license_link: LICENSE.md

configs:
- config_name: metadata
  data_files:
  - split: train
    path: "AAS_Metadata.csv"
    
---

<p align="center">
  <img
    src="https://huggingface.co/datasets/Appleartsstudios/Motion_Dataset/resolve/main/AAS_Logo.png"
    alt="Apple Arts Studios Logo"
    width="300"
  >
</p>

# Apple Arts Studios Motion Capture Dataset

## Dataset Overview

The **Apple Arts Studios Motion Capture Dataset** is a professionally captured, full-body human-motion dataset designed for:

- Artificial intelligence
- Humanoid robotics
- Motion generation
- Animation
- Simulation
- Action recognition
- Human-motion research

This release contains **5 hours of originally captured motion data**.

Left–right mirrored versions of the original motions are also included, increasing the total distributed duration to **10 hours**.

> **Original captured motion:** 5 hours  
> **Mirrored motion augmentation:** 5 hours  
> **Total distributed motion duration:** 10 hours

The mirrored motions are derived from the original recordings and should be treated as **data augmentation**, not as independently captured performances.

The dataset covers locomotion and full-body movement across:

- Flat rubbered floors
- Ramps
- Stairs
- Object-interaction environments

It includes multiple:

- Movement speeds
- Directions
- Body positions
- Transitions
- Movement styles
- Performer variations

Each motion is provided in the following representations:

- **SOMA Uniform BVH at 30 FPS**
- **Unitree G1 CSV**
- **JSON metadata**
- **Original and mirrored variants**
- **Structured environment, category, action, and actor hierarchy**

## Dataset Origin

This dataset, published by **Apple Arts Studios**, was conceived, performed, captured, processed, and curated entirely in-house at our own motion-capture facility for internal research and dataset development.

## Data Tiers

Apple Arts Studios motion data is available in two tiers. This Hugging Face repository provides a public evaluation release from Tier 1.

### Tier 1 — Segmented Motion Clips

Individually trimmed, cleaned, and training-ready clips. Each file contains one clearly defined action, separated by performer and variation, with T-poses, idle periods, and unrelated motion removed.

- Typical clip duration: **10–20 seconds**
- Formats: **SOMA Uniform BVH at 30 FPS** and **Unitree G1 CSV**
- Metadata: **Structured CSV and JSONL**
- Current release: **5 hours original + 5 hours mirrored = 10 hours**
- Full Tier 1 library:
  - Original: **67 hours 04 minutes**
  - Mirrored: **67 hours 04 minutes**
  - Total: **134 hours 08 minutes**

### Tier 2 — Full-Length Capture Sequences

Continuous motion-capture recordings that preserve the complete capture timeline, including T-poses, idle actions, transitions, resets, performer repositioning, and settle time between variations.

- Original: **195 hours 48 minutes**
- Mirrored: **195 hours 48 minutes**
- Total: **391 hours 36 minutes**

Synchronized witness-camera footage and OptiTrack Motive reference recordings are also available for the original capture sessions.

> **Note:** Tier 1 clips are segmented from the Tier 2 recordings and are not additional capture hours. Mirrored files are left-right augmented versions of the original motion and are not separately captured performances. Mirroring applies only to motion data, not to reference recordings.

## Licensing and Access

This Hugging Face release is provided as a public evaluation sample. The complete Tier 1 library, Tier 2 full-length sequences, additional action categories, and custom motion-capture services are available under a commercial licence.

For licensing and dataset access:

- **Email:** hello@appleartsstudios.com
- **Website:** https://www.appleartsstudios.com/

---

## Dataset Information

| Property | Description |
|---|---|
| **Dataset Provider** | Apple Arts Studios |
| **Dataset Type** | Full-body + fingers human motion capture |
| **Original Capture Duration** | 5 hours |
| **Mirrored Duration** | 5 hours |
| **Total Distributed Duration** | 10 hours |
| **Total Files** | 2,584 original and mirrored files combined |
| **Retargeted BVH Representation** | SOMA Uniform BVH |
| **SOMA Uniform Frame Rate** | 30 FPS |
| **Humanoid Robot Representation** | Unitree G1 |
| **Unitree G1 Format** | CSV |
| **Metadata Format** | JSON |
| **Motion Variants** | Original and mirrored |
| **Dataset Organization** | Environment, category, action, and actor |
| **Primary Applications** | AI, humanoid robotics, animation, simulation, and motion research |

---

# Motion Representations


## 1. SOMA Uniform BVH

The **SOMA Uniform BVH** files contain the same motions converted to the standardized SOMA Uniform skeletal representation.

These files are provided at **30 FPS** and use a consistent:

- Skeleton hierarchy
- Joint naming convention
- Bone structure
- Body proportion
- Reference representation

The standardized representation allows motions from different performers to be compared and processed using the same skeletal structure.

The SOMA Uniform representation may support:

- Motion-generation research
- Motion retrieval
- Skeletal representation learning
- Cross-actor motion comparison
- Standardized animation processing
- Human-motion machine learning
- Motion classification
- Motion prediction
- Motion embedding research

---

## 2. Unitree G1 CSV

The **Unitree G1 CSV** files provide motion data mapped to the Unitree G1 humanoid robot representation.

These files are intended for robotics research and may support:

- Human-to-robot motion retargeting
- Humanoid imitation learning
- Motion-policy development
- Simulation demonstrations
- Whole-body control research
- Physical-AI applications
- Reinforcement-learning research
- Robot-motion visualization

### Important Safety and Compatibility Notice

The Unitree G1 CSV files should **not** be treated as immediately deployable robot-control commands.

---

## 3. JSON Metadata

Each motion is accompanied by structured **JSON metadata**.

The metadata may include:

- Motion identifier
- Motion name
- Action name
- Category
- Subcategory
- Actor identifier
- Take number
- Version number
- Motion representation
- Frame rate
- Frame count
- Duration
- Original or mirrored status
- Movement direction
- Movement speed
- Object or prop information
- Relative file path
- Capture-system information
- Skeleton information

The metadata connects the corresponding SOMA Uniform BVH, and Unitree G1 CSV representations of the same motion.

---

# Frame-Rate Information

The dataset contains representation-specific frame rates.

| Representation | Frame Rate |
|---|---:|
| **SOMA Uniform BVH** | 30 FPS |
| **Unitree G1 CSV** | Recorded in the corresponding file metadata |

Users should **not assume that all distributed representations use the same frame rate**.

When comparing or synchronizing SOMA Uniform BVH, and Unitree G1 CSV files, users should account for:

- Resampling
- Frame indexing
- Frame count
- Motion duration
- Timestamp alignment
- Start-frame alignment
- Representation-specific processing

The filenames remain consistent across representations so that matching files can be identified and paired automatically.

---

# Original and Mirrored Motions

The dataset contains both original and left–right mirrored motion files.

## Original Motions

Original motions are captured performances processed through the Apple Arts Studios motion-capture pipeline.

## Mirrored Motions

Mirrored motions are derived from the corresponding original motion by applying left–right motion transformation.


Mirrored files are identified in the metadata using the original or mirrored status.

A mirrored motion should remain connected to its original source motion through a corresponding source-motion identifier.

> Mirrored motions are included as augmented data and are not counted as separately captured performances.

---

# Dataset Organization

The dataset is organized using a structured hierarchy based on:

1. Motion representation
2. Motion environment
3. Original or mirrored status
4. Surface or interaction type
5. Action category
6. Action name
7. Actor identifier

Example structure:

```text
AAS_Motion_Dataset/
│
├── Soma_Bvh/
│   └── Locomotions/
│       ├── Original/
│       └── Mirrored/
│
├── Unitree_G1_Csv/
│   └── Locomotions/
│       ├── Original/
│       └── Mirrored/