---
license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - LIBERO
  - imitation-learning
  - robotics
pretty_name: LIBERO Spatial Image — LIBERO Task 0
---

# LIBERO Spatial Image — LIBERO Task 0

This is a single-task subset derived from
[`lerobot/libero_spatial_image`](https://huggingface.co/datasets/lerobot/libero_spatial_image).
Its name follows the canonical task order of the LIBERO benchmark used at
runtime by verl-vla.

The selected task is LIBERO Spatial `task_id=0`:

> Pick up the black bowl between the plate and the ramekin and place it on the plate.

In the upstream LeRobot dataset's local task table, the same task is stored as
`task_index=4`. The numeric IDs differ because the upstream dataset and LIBERO
benchmark use different task-table orderings.

## Dataset summary

- Format: LeRobot v3.0
- Robot: Panda
- Frequency: 10 FPS
- Episodes: 45
- Frames: 4,487
- Tasks: 1
- Canonical LIBERO task ID: 0
- Upstream dataset-local task index: 4
- Observations: agent-view image, wrist image, and 8-dimensional robot state
- Actions: 7-dimensional continuous action

## Derivation and validation

All upstream episodes whose task description exactly matches the LIBERO task-0
description were selected. Observation and action values were copied without
modification, while episode, frame, sample, and dataset-local task indices were
rebuilt into a contiguous single-task LeRobot dataset.

Images are losslessly encoded as PNG data. The first extracted episode was
checked for exact tensor equality against its source episode for both images,
state, and action. Global image mean and standard deviation were recomputed
directly from all 4,487 decoded frames because the installed LeRobot version's
per-episode image-statistics aggregation produced invalid near-zero standard
deviations during subset creation.

This is a derived subset, not an independently collected dataset. Refer to the
upstream dataset and the [LIBERO project](https://libero-project.github.io/) for
the original collection and benchmark details.

## Loading with LeRobot

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

dataset = LeRobotDataset("Miical/libero_spatial_image_task0")
```

## License and attribution

The upstream dataset is published under the Apache-2.0 license. This derived
subset retains that license. Please credit the upstream LeRobot dataset and
LIBERO when using it.