---
license: apache-2.0
task_categories:
- robotics
tags:
- robotics
- lerobot
- so101
- scrape-isaaclab
- fold-towel
- temporal-downsampling
- 10fps
- stride-9
---

# SO101 Cap Fold Towel 10fps Stride9

LeRobot v3.0 derivative of `CoRL2026-CSI/SO101-cap_fold_towel_10fps`.

- **Robot**: `so101_follower`
- **Episodes**: 100
- **Frames**: 24,575
- **Nominal FPS**: 10
- **Cameras**: `observation.images.top`, `observation.images.left_wrist`, `observation.images.right_wrist`
- **Action keys**: `action`, `action.left_radian_urdf0`, `action.right_radian_urdf0`

## Derivation

This dataset keeps every 9th frame from the source dataset while preserving the nominal LeRobot `fps=10` metadata. The effective source-time compression is 9x relative to the source dataset.

Source frame mapping within each episode:

```text
output_frame_index k -> source_frame_index k * 9
output_timestamp    -> k / 10
```

For each retained source frame `i`, all observation, image, state, action, and skill columns are copied from source frame `i`. In particular, action labels are not relabelled from future states and are not shifted to the last skipped frame. This preserves the source dataset's causal LeRobot row pairing for the retained frames.

## Contents

Required state/action columns are preserved:

- `observation.state` (12 DoF)
- `action` (12 DoF)
- `observation.state.left_radian_urdf0` (6 DoF)
- `observation.state.right_radian_urdf0` (6 DoF)
- `action.left_radian_urdf0` (6 DoF)
- `action.right_radian_urdf0` (6 DoF)

All three RGB video streams are preserved and re-encoded at 10 fps.

## Provenance

- **Source dataset**: `CoRL2026-CSI/SO101-cap_fold_towel_10fps`
- **Derived dataset**: `CoRL2026-CSI/SO101-cap_fold_towel_10fps_stride9`
- **Frame stride**: 9
- **Source episodes**: 0-99
- **Generated with**: `scripts/tools/dataset/derive_subset_fps.py`