File size: 4,513 Bytes
3e1cc47 d605915 3e1cc47 d605915 3e1cc47 d605915 3e1cc47 d605915 3e1cc47 d605915 3e1cc47 d605915 3e1cc47 d605915 3e1cc47 d605915 3e1cc47 d605915 3e1cc47 d605915 3e1cc47 d605915 3e1cc47 d605915 3e1cc47 d605915 3e1cc47 d605915 3e1cc47 d605915 3e1cc47 d605915 3e1cc47 d605915 | 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 | ---
pretty_name: StreamingWAM Real-World Dataset
language:
- zh
tags:
- robotics
- imitation-learning
- real-world
- streamingwam
- multi-view-video
size_categories:
- 10GB<n<100GB
---
# StreamingWAM Real-World Dataset
This repository contains the raw real-world robot trajectories used for the
StreamingWAM experiments. It contains 590 trajectories covering seven semantic
tasks. Static and dynamic apple-to-bowl conditions are sampled independently
with 50 trajectories each and stored as separate subsets, giving eight
top-level task-condition subsets.
## Dataset summary
- **Trajectories:** 590
- **Semantic tasks:** 7
- **Task-condition subsets:** 8
- **Control/video rate:** 15 Hz
- **Camera views:** 3 synchronized MP4 streams per trajectory
- **Videos:** 1,770 H.264 files
- **Action/state files:** one JSON and one HDF5 file per trajectory
- **Timestamp files:** one JSON file per trajectory
- **Payload size:** 18,688,099,319 bytes
| Directory | English description | Trajectories |
| --- | --- | ---: |
| `先称苹果，然后把它放进碗里（统一15Hz版）` | Weigh an apple, then place it in a bowl | 100 |
| `关上抽屉（统一15Hz版）` | Close a drawer | 80 |
| `打地鼠（统一15Hz版）` | Whack-a-mole | 30 |
| `打开抽屉（统一15Hz版）` | Open a drawer | 80 |
| `把苹果放进碗里_小车（统一15Hz版）` | Place an apple in a bowl with a moving cart | 100 |
| `把苹果放进碗里_静态（统一15Hz版）` | Place an apple in a bowl: static condition | 50 |
| `把苹果放进碗里_动态（统一15Hz版）` | Place an apple in a bowl: dynamic condition | 50 |
| `搭积木` | Stack blocks | 100 |
## Record layout
Each trajectory contains one action JSON, one action HDF5, one frame-timestamp
JSON, and three synchronized camera MP4 files. The action data include robot
joint state, end-effector pose, gripper state, and timestamps. All streams are
aligned at 15 Hz. Most videos are 1920×1080; the moving-cart subset uses
1280×720 video.
## Selection and integrity
Static and dynamic apple-to-bowl records were sampled independently with 50
trajectories each. Other subsets retain the original publication policy of at
most 100 trajectories. Sampling is without replacement after sorting record
names and uses seed 42. The exact selection is recorded in
`SELECTION_MANIFEST.json`.
`SHA256SUMS.tsv` contains the SHA-256 digest, byte size, and repository-relative
path of all 3,540 payload files. Every payload file was hashed after the
final static/dynamic split. `DATASET_BUILD.md` and each subset's
`SOURCE_README.md` document the preparation procedure.
## Loading
```python
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="rsj2003/StreamingWAM-RealWorld",
repo_type="dataset",
local_dir="StreamingWAM-RealWorld",
)
```
This is a raw-data release. Use the 15 Hz timestamps when converting
trajectories to a training format, and do not apply a second 30 Hz-to-15 Hz
downsampling pass.
## Privacy and license
The videos were recorded in a laboratory robot workcell. Before making this
private repository public, confirm the intended release scope, institutional
consent requirements, and any incidental human appearance in the full video
collection. No dataset license has been selected yet; external components and
benchmark assets retain their respective terms.
## Data processing scripts
The public conversion and validation scripts are maintained in the [StreamingWAM code repository](https://github.com/renshaojie233/StreamingWAM/tree/main/data_processing/realworld_7task_15hz). They define the exact 15 Hz data contract used for the formal training conversion, including camera order, timestamp handling, state/action construction, and FastWAM tensor layout:
- [Conversion script](https://github.com/renshaojie233/StreamingWAM/blob/main/data_processing/realworld_7task_15hz/convert_realworld_15hz_multitask_to_lerobot_v21.py)
- [Validation script](https://github.com/renshaojie233/StreamingWAM/blob/main/data_processing/realworld_7task_15hz/validate_realworld_15hz_multitask_v21.py)
- [Processing README and checksums](https://github.com/renshaojie233/StreamingWAM/tree/main/data_processing/realworld_7task_15hz)
The scripts are read-only with respect to the source recordings and write to a fresh output directory. The public 590-trajectory release is a publication subset; the full formal training source contains additional episodes, as documented in the processing README.
|