ivlabs/so101_pick_cube_v2_merged_rlds

---
license: apache-2.0
tags:
- robotics
- lerobot
- vla
- rlds
- so-101
- manipulation
language:
- en
pretty_name: SO101 Pick-Cube RLDS (train/val)
size_categories:
- 10K<n<100K
---

# SO101 Pick Cube — RLDS TFRecords (train/val)

RLDS-formatted TFRecord dataset derived from [`sapanostic/so101_pick_cube_v2_merged`](https://huggingface.co/datasets/sapanostic/so101_pick_cube_v2_merged).
Produced by the `rebis-vla` conversion pipeline with a **90/10 train/val split** (shuffle-based, seed=42).

## Contents

| Split | Episodes | Size |
|---|---|---|
| `train` | 90 | 1067 MB |
| `val` | 10 | 150 MB |

Source: 100 episodes @ 15 Hz, SO-100/SO-101 6-DOF arm, task `"pick up the cube"`.

## Schema (per step)

Both `observation/` (measured) and `action/` (commanded) namespaces carry matching joint-space and EE-space representations.

| Field | Shape | Description |
|---|---|---|
| `observation/joint_angles_deg` | (6,) float32 | Measured joints in degrees |
| `observation/eef_pose_m_rad` | (7,) float32 | Measured EE pose `[x, y, z, roll, pitch, yaw, gripper_norm]` |
| `observation/eef_delta_m_rad` | (7,) float32 | Measured EE delta vs previous step (first = zeros) |
| `action/joint_angles_deg` | (6,) float32 | Commanded joint targets in degrees |
| `action/eef_pose_m_rad` | (7,) float32 | FK of commanded joints |
| `action/eef_delta_m_rad` | (7,) float32 | Commanded EE delta (first = zeros) |
| `observation/image_left` | (256, 256, 3) uint8 | Left camera RGB |
| `observation/image_right` | (256, 256, 3) uint8 | Right camera RGB |
| `observation/image_wrist` | (256, 256, 3) uint8 | Wrist camera RGB |
| `language_instruction` | string | "pick up the cube" |

Gripper normalized to [0, 1] against SO101 URDF limits (-10 deg to 100 deg), binarized at 0.5 by default. `joint_angles_deg` keeps gripper in raw degrees; EE pose/delta use the normalized value.

## Loading

```python
from huggingface_hub import snapshot_download
import tensorflow_datasets as tfds

root = snapshot_download("ivlabs/so101_pick_cube_v2_merged_rlds", repo_type="dataset")

train = tfds.builder_from_directory(f"{root}/train").as_dataset(split="train")
val = tfds.builder_from_directory(f"{root}/val").as_dataset(split="train")
```

## Provenance

- Source: [`sapanostic/so101_pick_cube_v2_merged`](https://huggingface.co/datasets/sapanostic/so101_pick_cube_v2_merged) (LeRobot, 100 ep @ 15 Hz)
- Robot: SO-100 / SO-101 6-DOF arm
- Conversion: `rebis-vla` pipeline (`convert_hf_to_rlds.py` + `tfrecord_to_tfds.py`)
- Split: 90/10 train/val, shuffle-based, seed=42. Val episodes: `[18, 21, 27, 33, 42, 50, 56, 59, 71, 92]`
- Images: resized 480x640 -> 256x256, JPEG quality 95
- FK: Pinocchio on SO101 URDF (`so101_new_calib.urdf`)