README.md

---
license: mit
task_categories:
- robotics
tags:
- LeRobot
- robotics
- preference-learning
- reward-modelling
- rlhf
- manipulation
- franka
configs:
- config_name: default
  data_files: data/*/*.parquet
---

# fold_pants — pairwise preferences on a Franka Panda

Real-robot trajectories for **"fold the shorts"** with **human pairwise preference
labels on multiple judgment axes**. Built for reward-model / preference-learning
research: every label is a comparison of two trajectories on one named axis, not a
scalar score.

The trajectory data is a standard [LeRobot](https://github.com/huggingface/lerobot)
v2.1 dataset, so it also loads directly as an imitation-learning dataset.

## Contents

| | |
|---|---|
| Episodes | **536** |
| Frames | **530,494** (~9.8 h at 15 fps) |
| Preference pairs | **1448** |
| Preference labels (pair x axis) | **4390** |
| Judgment axes | **45** |
| Distinct instructions | **1** |
| Cameras | `agent_view` (third-person), `wrist` — both 224x224 |
| Robot | Franka Panda, 7-DoF joint control + gripper |

### Episodes by kind

| kind | episodes | frames |
|---|---|---|
| rollout | 399 | 385498 |
| demo | 137 | 144996 |

`demo` episodes are human teleoperated demonstrations. `rollout` episodes are policy
rollouts recorded in preference-collection sessions; those are the ones that vary in
quality, which is what makes the comparisons informative.

## Preference labels

`preferences/pairs.parquet` — one row per (pair, axis):

| column | meaning |
|---|---|
| `pair_id` | unique id of the comparison |
| `episode_index_a`, `episode_index_b` | index into this LeRobot dataset |
| `episode_id_a`, `episode_id_b` | original source trajectory id |
| `axis` | the judgment axis being compared |
| `winner` | `A`, `B`, or `Equal` |
| `source` | `in_session` (the two rollouts of one session) or `cross` (arbitrary pair) |
| `source_dir`, `source_file` | provenance of the annotation |
| `annotator` | which annotator's label set this row comes from |
| `axis_set` | `fixed` (curated rubric reused across sessions) or `freeform` (annotator-invented axis names) |
| `overall_score_a/b` | 1–4 Likert quality rating, where the annotator gave one (else null) |
| `succeeded_a/b` | task-success flag recorded at collection time, where available (else null) |
| `instruction` | instruction string logged with the comparison |

`preferences/pairs.jsonl` is the same data with the axes nested per pair.
`preferences/episodes.parquet` maps `episode_index` to the source trajectory.

Winner distribution: **A** 1996, **B** 1968, **Equal** 426.

### Axes

| axis | labels |
|---|---|
| `Overall quality` | 1330 |
| `Quality of 1st fold` | 239 |
| `Quality of 2nd fold` | 239 |
| `Wrinkle of 2nd fold` | 239 |
| `Alignment of final fold` | 239 |
| `Fast` | 239 |
| `Smooth` | 239 |
| `Wrinkle of 1st fold` | 239 |
| `Damage to environment` | 238 |
| `Quality of 3rd fold` | 237 |
| `Wrinkle of 3rd fold` | 237 |
| `fast` | 69 |
| `wrinkle of first fold` | 48 |
| `speed` | 47 |
| `alignment of final fold` | 46 |
| `wrinkle of second fold` | 42 |
| `first fold quality` | 39 |
| `third fold quality` | 38 |
| `smoothness` | 32 |
| `second fold quality` | 30 |
| `final alignment` | 28 |
| `smooth` | 28 |
| `stable` | 26 |
| `wrinkle of third fold` | 26 |
| `first fold wrinkle` | 19 |
| `final fold alignment` | 19 |
| `quality of second fold` | 18 |
| `stability` | 16 |
| `second fold wrinkle` | 16 |
| `third fold wrinkle` | 16 |
| `quality of third fold` | 14 |
| `stablity` | 12 |
| `environment damage` | 11 |
| `quality of first fold` | 10 |
| `damage to environment` | 10 |
| `destruction to environment` | 5 |
| `environmental damage` | 2 |
| `damage caused to environment` | 1 |
| `Quality of placement of big plate` | 1 |
| `Quality of placement of small plate` | 1 |
| `Quality of placement of cup` | 1 |
| `Quality of placement of cutlery` | 1 |
| `Smoothness / carefulness` | 1 |
| `Speed` | 1 |
| `Formality of setup` | 1 |

### Pairs by source

| source | directory | axis set | pairs |
|---|---|---|---|
| cross | `abhijnya/cross_preferences` | fixed | 61 |
| cross | `am208/cross_preferences` | fixed | 690 |
| cross | `am208/cross_preferences_extra` | fixed | 498 |
| in_session | `am208/preferences` | fixed | 79 |
| in_session | `am208/preferences_fold_pants_free` | freeform | 120 |

## Usage

```python
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("MarcelTorne/fold_pants_preferences")
print(ds[0].keys())
```

```python
# train a reward model on the pairwise labels
import pandas as pd
from huggingface_hub import hf_hub_download
prefs = pd.read_parquet(hf_hub_download("MarcelTorne/fold_pants_preferences", "preferences/pairs.parquet",
                                        repo_type="dataset"))
decisive = prefs[prefs.winner != "Equal"]           # drop ties
overall  = decisive[decisive.axis.str.lower().str.contains("overall")]
```

Both cameras are `dtype: video`. `observation.state` is
`[joint_0..joint_6, gripper]`, `observation.ee` is `[x, y, z, roll, pitch, yaw]`,
and `action` is the commanded `[joint_position(7), gripper_position(1)]`.

## Caveats

- **Frames are h264 (crf 16), not bit-exact.** The source HDF5 held raw uint8
  frames; re-encoding at the same 224x224 resolution shrinks the release ~100x at
  41.8 dB PSNR (mean absolute error 1.1/255, 99th pct 7/255), measured on
  fold_pants cloth texture. Do not expect byte-level reproduction of the originals.
- **`Equal` is common on some axes.** Annotators used it freely; filter deliberately.
- **`freeform` axis names are not a closed vocabulary.** They were invented per
  session, so the same idea appears under several names (`fast` / `speed`,
  `smooth` / `smoothness`). Normalise before aggregating across sessions.
- **A pair can be labelled by more than one annotator.** The two annotators' cross-pair
  sets are near-disjoint, but not perfectly; group by `pair_id` if you need unique pairs.
- **Instructions on `rollout` episodes sometimes carry reward conditioning** (e.g.
  `"fold the shorts, fast: 1.0, ..."`) because the policy that produced them was
  reward-conditioned. The LeRobot task string is the plain instruction; the logged
  string is kept in `preferences/episodes.parquet`.
- **No success labels on `demo` episodes** — they are demonstrations, assumed good.

## Related datasets

Same robot, same collection pipeline, same schema — four tasks released together:

- **`MarcelTorne/fold_pants_preferences`** (fold the shorts) — this dataset
- [`MarcelTorne/setup_table_preferences`](https://huggingface.co/datasets/MarcelTorne/setup_table_preferences) — set up the table
- [`MarcelTorne/put_cube_in_bowl_preferences`](https://huggingface.co/datasets/MarcelTorne/put_cube_in_bowl_preferences) — put the cube in the bowl
- [`MarcelTorne/plate_toast_preferences`](https://huggingface.co/datasets/MarcelTorne/plate_toast_preferences) — put the toast in the plate

## License

MIT.