README.md

---
license: other
task_categories:
  - robotics
tags:
  - robotics
  - rlds
  - tfds
  - rh20t
  - franka
pretty_name: RH20T cfg5 Franka RLDS
---

# RH20T cfg5 Franka RLDS

This dataset is an RLDS/TFDS conversion of the RH20T Franka subset. It was
converted from RH20T `cfg5`, which corresponds to the Franka robot
configuration, and is intended to be loaded as the `rh20t` dataset in the
VQ-VLA/Open-X RLDS pipeline.

## Contents

The dataset is stored in TFDS directory format:

```text
rh20t/
  README.md
  1.0.0/
    dataset_info.json
    features.json
    dataset_statistics_*.json
    rh20t-train.tfrecord-00000-of-00008
    ...
    rh20t-train.tfrecord-00007-of-00008
    skipped_episodes.txt
```

## Conversion Summary

- Source dataset: RH20T
- Source configuration: `cfg5`
- Robot: Franka
- Output format: RLDS/TFDS TFRecord
- Number of trajectories: 1143
- Number of transitions: 660128
- Number of TFRecord shards: 8
- Skipped episodes: 88, recorded in `1.0.0/skipped_episodes.txt`

## Observation Space

Each step contains:

- `observation/image_front`: external front RGB image
- `observation/image_side_right`: external side-right RGB image
- `observation/image_wrist`: wrist RGB image
- `observation/proprio`: 7D proprioceptive state

The camera mapping used during conversion is:

- `image_front`: camera `037522062165`
- `image_side_right`: camera `104122061850`
- `image_wrist`: camera `135122079702`, with fallback to `104422070042`

The proprio vector is:

```text
[eef_x, eef_y, eef_z, eef_roll, eef_pitch, eef_yaw, gripper_width]
```

## Action Space

The action vector is a 7D end-effector action:

```text
[delta_x, delta_y, delta_z, delta_roll, delta_pitch, delta_yaw, gripper]
```

The first six dimensions are relative end-effector pose deltas in the robot base
frame. The final dimension is an absolute gripper command:

- `1.0`: open
- `0.0`: close

This matches the `ActionEncoding.EEF_POS` convention in the VQ-VLA Open-X
configuration for `rh20t`.

## VQ-VLA/OXE Config

This dataset is designed to match the following dataset config:

```python
"rh20t": {
    "image_obs_keys": {
        "primary": "image_front",
        "secondary": "image_side_right",
        "wrist": "image_wrist",
    },
    "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
    "state_obs_keys": ["proprio"],
    "state_encoding": StateEncoding.POS_EULER,
    "action_encoding": ActionEncoding.EEF_POS,
}
```

## Verification

The converted dataset was checked with:

- Full TFRecord shard scan: all 8 shards were readable.
- TFDS builder load from `rh20t/1.0.0`.
- VQ-VLA Open-X RLDS loader using `primary,secondary,wrist` camera views.

Verified loaded shapes with a future action window of 6:

```text
image_primary:   (218, 1, 224, 224, 3)
image_secondary: (218, 1, 224, 224, 3)
image_wrist:     (218, 1, 224, 224, 3)
action:          (218, 7, 7)
proprio:         (218, 1, 7)
```

## Loading

Use this dataset as a local TFDS directory under an Open-X data root, e.g.:

```text
DATA_ROOT_DIR=/path/to/OXE
PRETRAIN_DATASET_NAME=rh20t
LOAD_CAMERA_VIEWS=primary,secondary,wrist
```

## Notes

This is not the full RH20T dataset. It only contains the Franka `cfg5` subset
converted to RLDS format.