---
license: cc-by-4.0
pretty_name: T-Rex Dataset Visualizer
task_categories:
- robotics
- video-classification
language:
- en
tags:
- robotics
- tactile-sensing
- dexterous-manipulation
- bimanual-manipulation
- teleoperation
- video
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: metadata.csv
---

# T-Rex Dataset Visualizer

A browseable subset of the **T-Rex** dataset — *Tactile-Rich Bimanual Dexterous
Manipulation* — collected on a bimanual **Dexmate Vega-1** robot equipped with
two **Sharpa Wave** dexterous hands.

This visualizer subset contains **3,838 short trajectory clips** drawn from the
full 100-hour T-Rex collection, organized by `(verb, object, hand)` so you can
quickly inspect coverage across motion primitives and object categories.

> For the full dataset (multi-view RGB, robot states, joint actions, per-fingertip
> tactile signals, and language instructions), see the main project release.

## Schema

| Column        | Type   | Description                                                     |
|---------------|--------|-----------------------------------------------------------------|
| `file_name`   | video  | Relative path to the `.mp4` clip (head camera view).            |
| `id`          | string | Unique 5-digit clip identifier.                                 |
| `verb`        | string | Canonical motion primitive (e.g. `reach`, `grasp_and_lifting`). |
| `object`      | string | Canonical object category (e.g. `book`, `airpods`).             |
| `hand`        | string | `left_hand` or `right_hand`.                                    |
| `episode`     | string | Episode identifier the clip was sampled from.                   |
| `verb_raw`    | string | Raw verb label before canonicalization.                         |
| `object_raw`  | string | Raw object label before canonicalization.                       |

## Coverage

- **Total clips:** 3,838
- **Motion primitives:** 22 (e.g. `grasp_and_lifting`, `reach`, `wrap`, `fold`,
  `press`, `peel`, `wipe`, `squeeze`, `open`, `cut`, `close`, `shake`, `pour`,
  `screw`, `insert`, `assemble`, …)
- **Object categories:** 270+ daily objects (top: `airpods`, `bubble_wrap`,
  `book`, `aluminum_foil`, `button`, `clamshell_container`, `flip_phone`,
  `bowl`, `sponge`, `plastic_wrap`, …)
- **Hands:** balanced left / right coverage

A full breakdown is provided in [`manifest.json`](./manifest.json).

## Usage

```python
from datasets import load_dataset

ds = load_dataset("Beakerman0101/trex-visualizer", split="train")
print(ds[0])
# {'file_name': <Video>, 'id': '00001', 'verb': 'reach', 'object': 'book',
#  'hand': 'left_hand', 'episode': 'episode_0000', ...}
```

You can also stream raw clips directly by URL:

```
https://huggingface.co/datasets/Beakerman0101/trex-visualizer/resolve/main/videos/<file_name>
```

## Filtering Examples

```python
# All "fold" demonstrations
folds = ds.filter(lambda r: r["verb"] == "fold")

# Left-hand "grasp_and_lifting" of a book
subset = ds.filter(
    lambda r: r["verb"] == "grasp_and_lifting"
              and r["object"] == "book"
              and r["hand"] == "left_hand"
)
```

## License

Released under **CC-BY-4.0**. Please cite the T-Rex project if you use this
visualizer subset in published work.