djkesu/tshirt-folding

---
pretty_name: T-Shirt Folding Mixed Manifest
tags:
- robotics
- imitation-learning
- lerobot
- openpi
task_categories:
- other
size_categories:
- n<1K
---

# T-Shirt Folding Mixed Manifest

This repo is a lightweight manifest for a public T-shirt folding collection built from five source datasets:

- `Gongsta/trlc_tshirt_folding`
- `Gongsta/trlc_tshirt_folding_impedance`
- `Gongsta/dagger_dk1_tshirt_corrections`
- `Gongsta/krish-simpler-tshirt`
- `Gongsta/e7-tshirt-folding`

The goal is to provide one clean public entrypoint while preserving each source dataset at its highest native frame rate.

## Project Background

This dataset card was created from data collected across Waterloo for the University of Waterloo Software Engineering capstone by team members Eddy Zhou, Steven Gong, Krish Shah, and Krish Mehta. We used these datasets to train a laundry-folding robot by fine-tuning the base Pi-0.5 model, including runs that incorporated DAgger corrections.

In practice, we saw strong qualitative accuracy and reasonable generalization across two environments and different T-shirts. We were not highly rigorous about exact benchmark evaluation, so this collection should be treated more as a practical public starting point than as a tightly standardized benchmark. The hope is that it helps other people train stronger models with more careful evaluation and higher final precision.

## Best Observed Recipe

In our own experiments, the best model quality we observed came from:

- training at `50 Hz`
- using impedance-control data
- `batch_size=32`
- `5000` training steps
- `4x A100`
- training on `Gongsta/e7-tshirt-folding`

That result should be treated as an empirical recipe from our runs, not a universal rule.

## What Is In This Repo

Each row in `data/train.jsonl` describes one source dataset and includes:

- `source_repo`
- `source_tag`
- `native_fps`
- `is_dagger`
- `is_impedance`
- `scene_type`
- `shirt_layout`
- `recommended_common_fps`

This repo does **not** duplicate the source videos. It is a manifest / collection layer that documents provenance and intended loading behavior.

## Source Tags

- `apartment_original_multi_shirt`
- `apartment_impedance_multi_shirt`
- `multi_shirt_dagger_corrections`
- `apartment_single_shirt`
- `building_single_shirt`

## Recommended Loading

If you want to combine all five datasets with exact-stride downsampling and no interpolation, use:

- `target_fps = 10`

Why:

- `30 -> 10` is exact stride `3`
- `50 -> 10` is exact stride `5`

If you want to keep native FPS, load the listed source repos directly and treat this dataset as the metadata / provenance index.

## Using Standard LeRobot Loaders

This repo is intended for users who may only have the standard LeRobot dataset tools, not our internal training code.

The simplest pattern is:

1. read `data/train.jsonl` from this manifest repo
2. choose the source repos you want
3. load those source repos directly with `LeRobotDataset`
4. decide whether to keep native FPS or resample to a common target FPS yourself

If you want a common timebase across all five datasets, our recommendation is:

- downsample everything to `10 Hz`
- use exact stride downsampling where possible

Recommended exact-stride logic:

- `30 Hz -> 10 Hz`: keep every `3rd` frame/state/action
- `50 Hz -> 10 Hz`: keep every `5th` frame/state/action

This avoids interpolation entirely for the current dataset set.

If you want to upsample instead, we recommend:

- interpolate proprioceptive signals such as state and action sequences
- use nearest-frame selection for images rather than inventing intermediate video frames

For most users, exact-stride downsampling is the simpler and more reproducible choice.

## Notes

- `Gongsta/trlc_tshirt_folding` is the only `30 Hz` source.
- `Gongsta/dagger_dk1_tshirt_corrections` is the DAgger dataset and is `50 Hz`.
- The other listed sources are `50 Hz`.

## Example: Load Through This Manifest

```python
from datasets import load_dataset
from lerobot.datasets.lerobot_dataset import LeRobotDataset

manifest = load_dataset("djkesu/tshirt-folding", split="train")

# Example: choose all native-50Hz, non-dagger datasets
selected = manifest.filter(
lambda row: row["native_fps"] == 50 and not row["is_dagger"]
)

repo_ids = [row["source_repo"] for row in selected]
datasets = [LeRobotDataset(repo_id=repo_id) for repo_id in repo_ids]
```

## Example: Use All Listed Source Datasets

```python
from datasets import load_dataset

manifest = load_dataset("djkesu/tshirt-folding", split="train")
repo_ids = [row["source_repo"] for row in manifest]

# For an exact-stride common timebase across all five:
target_fps = 10
```