RGB-D hand manipulation dataset captured with iPhone 13 TrueDepth sensor for humanoid robot training. Includes 6-DoF hand pose trajectories, synchronized video, and semantic motion annotations.

📊 Dataset Overview

Metric	Value
Episodes	97
Total Frames	~28,000
FPS	30
Tasks	10 manipulation tasks
Total Duration	~15.5 minutes
Avg Episode Length	~9.6 seconds

Task Distribution

Task ID	Description	Episodes
Task 1	Fold the white t-shirt on the bed	8
Task 2	Fold the jeans on the bed	10
Task 3	Fold two underwear and stack them	10
Task 4	Put the pillow on the right place	10
Task 5	Pick up plate and glass, put on stove	10
Task 6	Go out the door and close it	9
Task 7	Pick up sandals, put next to scale	10
Task 8	Put cloth in basket, close drawer	10
Task 9	Screw the cap on your bottle	10
Task 10	Pick up two objects, put on bed	10

📁 Repository Structure

humanoid-robots-training-dataset/
│
├── data/
│   └── chunk-000/                    # Parquet files (97 episodes)
│       ├── episode_000000.parquet
│       ├── episode_000001.parquet
│       └── ...
│
├── videos/
│   └── chunk-000/rgb/                # MP4 videos (synchronized)
│       ├── episode_000000.mp4
│       └── ...
│
├── meta/                             # Metadata & Annotations
│   ├── info.json                     # Dataset configuration (LeRobot format)
│   ├── stats.json                    # Feature min/max/mean/std statistics
│   ├── events.json                   # Disturbance & recovery annotations
---

## 🎯 Data Schema

### Parquet Columns (per frame)

| Column | Type | Description |
|--------|------|-------------|
| `episode_index` | int64 | Episode number (0-96) |
| `frame_index` | int64 | Frame within episode |
| `timestamp` | float64 | Time in seconds |
| `language_instruction` | string | Task description |
| `observation.state` | float[252] | 21 hand joints × 2 hands × 6 DoF |
| `action` | float[252] | Same as state (for imitation learning) |
| `observation.images.rgb` | struct | Video path + timestamp |

### 6-DoF Hand Pose Format

Each joint has 6 values: `[x_cm, y_cm, z_cm, yaw_deg, pitch_deg, roll_deg]`

**Coordinate System:**
- Origin: Camera (iPhone TrueDepth)
- X: Right (positive)
- Y: Down (positive)  
- Z: Forward (positive, into scene)

---

## 🏷️ Motion Semantics Annotations

**File:** `meta/annotations_motion_v1_frames.json`

Coarse temporal segmentation with motion intent, phase, and error labels.

### Annotation Schema

```json
{
  "episode_id": "Task1_Vid2",
  "segments": [
    {
      "start_frame": 54,
      "end_frame_exclusive": 140,
      "motion_type": "grasp",           // What action is being performed
      "temporal_phase": "start",        // start | contact | manipulate | end
      "actor": "both_hands",            // left_hand | right_hand | both_hands
      "target": {
        "type": "cloth_region",         // cloth_region | object | surface
        "value": "bottom_edge"          // Specific target identifier
      },
      "state": {
        "stage": "unfolded",            // Task-specific state
        "flatness": "wrinkled",         // For folding tasks only
        "symmetry": "asymmetric"        // For folding tasks only
      },
      "error": "none"                   // misalignment | slip | drop | none
    }
  ]
}

Motion Types

Why Motion Annotations?

Temporal Structure: Know when manipulation phases begin/end
Intent Understanding: What the human intends to do, not just kinematics
Error Detection: Labeled failure modes (slip, drop, misalignment)
Training Signal: Richer supervision for imitation learning

📋 Events Metadata

File: meta/events.json

Disturbances and recovery actions for select episodes.

Disturbance Types

Type	Description
`OCCLUSION`	Hand temporarily blocked from camera
`TARGET_MOVED`	Object shifted unexpectedly
`SLIP`	Object slipped during grasp
`COLLISION`	Unintended contact
`DEPTH_DROPOUT`	Depth sensor lost valid readings

Recovery Actions

Action	Description
`REGRASP`	Release and re-acquire object
`REACH_ADJUST`	Modify approach trajectory
`ABORT`	Stop current action
`REPLAN`	Compute new action sequence

📈 Depth Quality Metrics

Metric	Description	Dataset Average
`valid_depth_pct`	% frames with valid depth at hand	95.5% ✅

🚀 Usage

With LeRobot

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("DynamicIntelligence/humanoid-robots-training-dataset")

# Access episode
episode = dataset[0]
state = episode["observation.state"]      # [252] hand pose (both hands)
rgb = episode["observation.images.rgb"]   # Video frame
task = episode["language_instruction"]    # Task description

Loading Motion Annotations

import json
from huggingface_hub import hf_hub_download

# Download annotations
path = hf_hub_download(
    repo_id="DynamicIntelligence/humanoid-robots-training-dataset",
    filename="meta/annotations_motion_v1_frames.json",
    repo_type="dataset"
)

with open(path) as f:
    annotations = json.load(f)

# Get segments for Task1
task1_episodes = annotations["tasks"]["Task1"]["episodes"]
for ep in task1_episodes:
    print(f"{ep['episode_id']}: {len(ep['segments'])} segments")

Combining Pose + Annotations

# Get frame-level motion labels
def get_motion_label(frame_idx, segments):
    for seg in segments:
        if seg["start_frame"] <= frame_idx < seg["end_frame_exclusive"]:
            return seg["motion_type"], seg["temporal_phase"]
    return None, None

# Example: label each frame
for frame_idx in range(episode["frame_index"].max()):
    motion, phase = get_motion_label(frame_idx, episode_annotations["segments"])
    if motion:
        print(f"Frame {frame_idx}: {motion} ({phase})")

📖 Citation

If you use this dataset in your research, please cite:

@dataset{dynamic_intelligence_2024,
  author = {Dynamic Intelligence},
  title = {Egocentric Human Motion Annotation Dataset},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/DynamicIntelligence/humanoid-robots-training-dataset}
}

📧 Contact

Email: shayan@dynamicintelligence.company
Organization: Dynamic Intelligence

🖼️ Hand Landmark Reference

Each hand has 21 tracked joints. The observation.state contains 6-DoF (x, y, z, yaw, pitch, roll) for each joint.

👁️ Visualizer Tips

When using the DI Hand Pose Sample Dataset Viewer:

Enable plots: Click the white checkbox next to joint names (e.g., left_thumb_cmc_yaw_deg) to show that data in the graph
Why not all enabled by default?: To prevent browser lag, only a few plots are active initially
Full data access: All joint data is available in the parquet files under Files and versions

Downloads last month: 463

You need to agree to share your contact information to access this dataset

Dynamic Intelligence - Egocentric Human Motion Annotation Dataset