README.md

---
license: apache-2.0
task_categories:
- robotics
- imitation-learning
tags:
- rlds
- tfds
- open-x-embodiment
- vlabench
- lerobot
- franka
- robot-learning
pretty_name: VLABench RLDS Delta EEF
---

# VLABench RLDS Delta EEF

This dataset is a TFDS/RLDS conversion of
[`VLABench/vlabench_composite_ft_lerobot_video`](https://huggingface.co/datasets/VLABench/vlabench_composite_ft_lerobot_video)
for the VQ-VLA / OXE RLDS training pipeline.

The converted action space is delta end-effector pose:

```text
[dx, dy, dz, droll, dpitch, dyaw, gripper_command]
```

## Source Dataset

- Source: `VLABench/vlabench_composite_ft_lerobot_video`
- Original format: LeRobot v3.0 parquet + MP4 videos
- Robot: Franka
- Episodes: 5,977
- Frames / transitions: 2,539,771
- Tasks: 167
- FPS: 10
- Camera views: front, secondary, wrist
- Image resolution: 224 x 224 RGB

## RLDS Layout

Use this dataset as a local TFDS directory by placing it under an OXE data root:

```text
/path/to/OXE/vlabench/1.0.0
```

The converted `steps` feature contains:

- `observation/image`: JPEG bytes, primary/front RGB view, 224 x 224
- `observation/second_image`: JPEG bytes, secondary RGB view, 224 x 224
- `observation/wrist_image`: JPEG bytes, wrist RGB view, 224 x 224
- `observation/EEF_state`: `float32[6]`, `[x, y, z, roll, pitch, yaw]`
- `observation/gripper_state`: `float32[1]`
- `observation/joint_state`: `float32[7]`
- `observation/state`: `float32[8]`, `[EEF_state, 0.0 padding, gripper_state]`
- `action`: `float32[7]`, delta EEF pose plus gripper command
- `language_instruction`: task instruction string
- `reward`, `discount`, `is_first`, `is_last`, `is_terminal`

## Action Conversion

The source LeRobot `actions` field stores absolute EEF target poses:

```text
[target_x, target_y, target_z, target_roll, target_pitch, target_yaw, gripper]
```

During conversion, each action is transformed into a delta pose relative to the
current EEF observation:

```text
delta_xyz = target_xyz - current_eef_xyz
delta_rpy = (R_target * R_current.inv()).as_euler("xyz")
gripper_command = clip(source_gripper, 0, 1)
```

`roll`, `pitch`, and `yaw` use XYZ Euler angles.

## Conversion Command

The dataset was generated with:

```bash
python scripts/data/vlabench/convert_vlabench_lerobot_to_rlds.py \
  --input-root /nvme/cuiluyi/resources/datasets/VLABench/vlabench_composite_ft_lerobot_video \
  --output-root /nvme/cuiluyi/resources/datasets/OXE \
  --num-shards 64 \
  --num-workers 8 \
  --jpeg-quality 3 \
  --ffmpeg-threads 1 \
  --overwrite
```

The converter decodes the source AV1 MP4 videos with ffmpeg/libdav1d and stores
per-frame JPEG bytes in TFRecord shards. It also scans the actual parquet files
to correct source metadata entries whose recorded data file pointer does not
match the local file layout.

## OXE Configuration

The corresponding OXE config is:

```python
"vlabench": {
    "image_obs_keys": {"primary": "image", "secondary": "second_image", "wrist": "wrist_image"},
    "depth_obs_keys": {"primary": None, "secondary": None, "wrist": None},
    "state_obs_keys": ["EEF_state", None, "gripper_state"],
    "state_encoding": StateEncoding.POS_EULER,
    "action_encoding": ActionEncoding.EEF_POS,
}
```

## Validation

The converted dataset was checked with:

- `tfds.builder_from_directory(...)` on `vlabench/1.0.0`
- Metadata checks: 5,977 trajectories, 2,539,771 transitions, 64 shards
- Source-to-converted sampled checks across multiple shards:
  - `max_action_diff == 0.0`
  - `max_eef_diff == 0.0`
  - `max_grip_diff == 0.0`
  - `max_state_diff == 0.0`
  - `max_joint_diff == 0.0`
- OXE `make_single_dataset(...)` smoke test with primary, secondary, and wrist
  images, proprio state, and 7D action loading successfully