README.md

---
license: cc-by-4.0
task_categories:
  - robotics
tags:
  - lerobot
  - tactile
  - manipulation
  - project-aria
  - egocentric
size_categories:
  - n<1K
---

# FEEL-Benchmark

A multimodal egocentric manipulation dataset combining **Project Aria** smart-glasses recordings (RGB video, eye gaze, hand tracking, IMU) with synchronized **Paxini** tactile-glove force readings. Released in [LeRobot v3.0](https://github.com/huggingface/lerobot) format (GR00T path variant).

## Overview

- **26 trials**, **700 episodes**, **~65.8 min** of segmented manipulation, ~2.0 GB
- **Recording dates**: 2026-01-10, 2026-01-21, 2026-01-28, 2026-04-30, 2026-05-06
- **Camera**: Project Aria RGB (1408×1408, **10 fps**, H.264)
- **Tactile**: bilateral force totals (`action.tactile.left.force_total`, `action.tactile.right.force_total`)
- **Hand pose**: Aria MPS hand tracking (left + right, validity flags, confidence)
- **Gaze**: Aria MPS general eye gaze (yaw/pitch + depth)
- **Mean episode length**: 5.6 s (39,497 frames @ 10 fps)

## Dataset statistics

### By recording date

| Date       | Trials | Episodes | Frames | Duration (min) |
|------------|-------:|---------:|-------:|---------------:|
| 2026-01-10 |      3 |       42 |  3,565 |            5.9 |
| 2026-01-21 |      7 |      137 | 11,106 |           18.5 |
| 2026-01-28 |      5 |      132 |  8,481 |           14.1 |
| 2026-04-30 |      2 |       82 |  4,761 |            7.9 |
| 2026-05-06 |      9 |      307 | 11,584 |           19.3 |
| **Total**  | **26** |  **700** | **39,497** |     **65.8** |

### By object category

| Category | Trials | Episodes | Frames | Duration (min) |
|----------|-------:|---------:|-------:|---------------:|
| Bottle   |      5 |      237 | 10,811 |           18.0 |
| Jelly    |      6 |      115 | 10,232 |           17.1 |
| Sponge   |      3 |      107 |  6,548 |           10.9 |
| Water    |      4 |       70 |  4,860 |            8.1 |
| Wangzai  |      4 |      116 |  4,294 |            7.2 |
| Coke     |      2 |       19 |  1,512 |            2.5 |
| Chip     |      2 |       36 |  1,240 |            2.1 |

The **Bottle** category includes a fill-level gradient (`Empty_bottle_1`, `P25_bottle`, `P50_bottle`, `P75_bottle`, `Full_bottle_1`) for studying tactile cues of liquid mass and slosh.

## Layout

Datasets are grouped by recording session date. Trial names encode the manipulation outcome where relevant (e.g. `Chip_intact` = potato chip not crushed, `Chip_crushed` = chip crushed during grasp).

```
FeelAuthors/FEEL-Benchmark/
├── 20260110/  Jelly_1, Jelly_2, Jelly_3                                                  (3 trials)
├── 20260121/  Half_coke_2, Half_water_1, Jelly_{5,6,7}, Quarter_coke_0, Quarter_water_2  (7 trials)
├── 20260128/  Half_water_{3,4}, Sponge_{1,2,3}                                           (5 trials)
├── 20260430/  Empty_bottle_1, Full_bottle_1                                              (2 trials)
└── 20260506/  Chip_intact, Chip_crushed, Empty_Want_Milk, Empty_wz, P25_bottle,
              P50_bottle, P75_bottle, Want_Milk, Wz                                       (9 trials)
```

Each `*_lerobot/` directory follows LeRobot v3.0:

```
<trial>_lerobot/
├── meta/
│   ├── info.json          # feature schema, fps, episode counts
│   ├── episodes.jsonl     # per-episode metadata
│   ├── stats.json
│   └── tasks.jsonl
├── data/chunk-000/episode_NNNNNN.parquet
└── videos/chunk-000/observation.images.main/episode_NNNNNN.mp4
```

## Feature schema

Each row in `episode_*.parquet`:

| Field                              | dtype       | shape | Notes                            |
|------------------------------------|-------------|-------|----------------------------------|
| `index`                            | int64       | (1,)  | Global frame index               |
| `episode_index`                    | int64       | (1,)  |                                  |
| `frame_index`                      | int64       | (1,)  | Per-episode frame index          |
| `timestamp` / `timestamp_ns`       | float64/int |       | Aria device clock                |
| `observation.state`                | float32     | (7,)  | Right-hand pose state            |
| `action`                           | float32     | (2,)  | Action vector                    |
| `status.hands.valid_right`         | bool        | (1,)  | Hand-tracking validity           |
| `action.tactile.left.force_total`  | float32     | (1,)  | Total force, left glove (N)      |
| `action.tactile.right.force_total` | float32     | (1,)  | Total force, right glove (N)     |
| `observation.images.main`          | video       | 1408×1408×3 | Aria RGB stream            |

## Path format note

Each `meta/info.json` declares `"path_format": "groot"`. The `data_path` and `video_path` templates match standard LeRobot v3.0 (`data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet`), so the dataset is consumable by both stock LeRobot tooling and NVIDIA Isaac GR00T loaders.

## Sensors

- **Project Aria** (RGB, eye-tracking cameras, SLAM cameras, IMUs) with MPS post-processing for gaze + hand tracking
- **Paxini tactile glove** for force feedback (totals exported per side; raw per-taxel data not included in this release)

## Loading

```python
from datasets import load_dataset
ds = load_dataset("FeelAuthors/FEEL-Benchmark",
                  data_files="20260506/P25_bottle_lerobot/data/chunk-000/*.parquet")
```

Or with LeRobot tooling, point a `LeRobotDataset` at a single trial directory.

## License

[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

## Citation

```bibtex
@misc{feel_benchmark_2026,
  title  = {FEEL-Benchmark: Egocentric Tactile-Visual Manipulation Dataset},
  author = {FeelAuthors},
  year   = {2026},
  url    = {https://huggingface.co/datasets/FeelAuthors/FEEL-Benchmark}
}
```