File size: 5,410 Bytes
0fd7c56
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
---
language:
  - en
task_categories:
  - robotics
  - reinforcement-learning
tags:
  - robotics
  - manipulation
  - imitation-learning
  - world-model
  - robot-learning
  - tabletop
pretty_name: "World Model Robot Manipulation Dataset (Our-50)"
size_categories:
  - n<1K
---

# World Model Robot Manipulation Dataset

A dataset of real-robot tabletop manipulation trajectories collected for world model training and imitation learning research. The setup follows DROID Dataset. Each trajectory pairs multi-camera video, proprioceptive state/action sequences, natural language task descriptions, and dense reward annotations with pre-extracted visual latents.

## Dataset Summary

| Split | Trajectories | Success Rate | Avg. Length |
|-------|-------------|--------------|-------------|
| Train | 250         | 44.8%        | 118 frames  |
| Val   | 100         | 44.0%        | 106 frames  |
| **Total** | **350** | **44.6%**  | **115 frames** |

Five tabletop manipulation tasks, 50 train / 20 val trajectories per task.

## Tasks

| Task ID | Description | Train SR | Val SR |
|---------|-------------|----------|--------|
| `bag_our`    | Pick up a bag of chips and place it on a green plate | 54% | 60% |
| `marker_our` | Pick up a marker and place it in a cup/mug           | 36% | 30% |
| `pour_our`   | Pick up a cup of beans and place them in a bowl      | 34% | 30% |
| `stack_our`  | Pick up a bowl and stack it on top of another bowl   | 60% | 60% |
| `towel_our`  | Pick up a towel and place it in a basket             | 40% | 40% |

Each task has multiple natural-language paraphrases (e.g. *"put the marker in the cup"*, *"put the marker in the mug"*, *"pick up the marker and place it in the cup"*).

## Data Structure

```
world_model_data_our_50/
├── annotations/
│   ├── train/   {0..249}.json
│   └── val/     {0..99}.json
├── annotation_rewards/
│   ├── train/   {0..249}.json       # same schema as annotations, includes reward fields
│   └── val/     {0..99}.json
├── latents/
│   ├── train/   {0..249}_sd3.npz
│   └── val/     {0..99}_sd3.npz
├── videos/
│   ├── train/   {0..249}.mp4
│   └── val/     {0..99}.mp4
├── norm_stats_recorded.json
└── norm_stats_relabel.json
```

### Annotation JSON Schema

Each `.json` file contains one trajectory with the following fields:

| Field | Type | Description |
|-------|------|-------------|
| `episode_id` | int | Sequential trajectory index within the split |
| `episode_id_orig` | str | Original episode identifier (e.g. `bag_our_003`) |
| `texts` | list[str] | Natural language task descriptions |
| `text_features` | float[768] | Pre-computed text embedding |
| `success` | int | Binary success label (1 = task completed) |
| `video_length` | int | Number of frames in the trajectory (32–334) |
| `video_path` | str | Relative path to the `.mp4` file |
| `latent_path` | str | Relative path to the latent `.npz` file |
| `num_cameras` | int | Always 3 |
| `states` | float[T][7] | Raw proprioceptive state per frame |
| `observation.state.cartesian_position` | float[T][6] | End-effector Cartesian pose (x, y, z, rx, ry, rz) |
| `observation.state.joint_position` | float[T][7] | 7-DOF joint positions |
| `observation.state.gripper_position` | float[T][1] | Gripper opening |
| `action.cartesian_position` | float[T][6] | Cartesian position action |
| `action.joint_position` | float[T][7] | Joint position action |
| `action.joint_velocity` | float[T][7] | Joint velocity action |
| `action.gripper_position` | float[T][1] | Gripper action |
| `reward_progress` | float[T] | Dense progress reward |
| `reward_success` | float[T] | Success-shaped reward |
| `reward_binary` | float[T] | Binary reward signal |

### Video Format

- Resolution: **960 × 192** (three 320 × 192 camera views (left, right, wrist) concatenated horizontally)
- Codec: H.264
- Frame rate: **5 fps**
- Length: 32–334 frames per trajectory

### Visual Latents

Pre-extracted with **Stable Diffusion 3** (SD3). Stored as `float16` NumPy arrays.

```
latents.npz  →  key: "latents"
shape: (3, T, 60, 256)
        │  │   │    └─ channel dim
        │  │   └─ spatial tokens
        │  └─ frames
        └─ cameras
```

### Normalization Statistics

`norm_stats_recorded.json` and `norm_stats_relabel.json` provide mean/std statistics for the `state` and `actions` modalities, suitable for normalizing inputs during training.

## Robot Setup

- **Robot**: Franka Emika Robot arm with parallel-jaw gripper (Robotiq Gripper)
- **Cameras**: 3 fixed cameras providing left, right, and wrist views
- **Control frequency**: 5 Hz (matches video frame rate)

## Usage Example

```python
import json
import numpy as np

# Load a trajectory
with open("annotations/train/0.json") as f:
    traj = json.load(f)

print(traj["texts"])          # ['pick up the bag of chips and place it on the green plate']
print(traj["success"])        # 1
print(traj["video_length"])   # e.g. 112

# Joint positions: shape (T, 7)
joint_pos = np.array(traj["observation.state.joint_position"])

# Actions: shape (T, 7)
actions = np.array(traj["action.joint_position"])

# Visual latents: shape (3, T, 60, 256)
lat = np.load(traj["latent_path"].replace("latents/", "latents/"))["latents"]

# Rewards: shape (T,)
rewards = np.array(traj["reward_progress"])
```