CollisionCode/libero_mem_lerobot_v2.1

---
license: mit
task_categories:
- robotics
tags:
- libero
- lerobot
- embodied-ai
- manipulation
- imitation-learning
pretty_name: LIBERO-Mem (LeRobot v2.1 Format)
size_categories:
- 10G<n<100G
---
# LIBERO-Mem (LeRobot v2.1 Format)

## Dataset Description
This dataset is a converted version of [**LIBERO-Mem**](https://huggingface.co/datasets/libero-mem/LIBERO-Mem), now formatted for compatibility with the **[LeRobot](https://github.com/huggingface/lerobot)** ecosystem (v3.0).

While the official LIBERO dataset only provides HDF5 files and RLDS conversion methods, this version enables seamless training with modern imitation learning pipelines like Diffusion Policy, ACT, and VLA models within the LeRobot framework.

- **Original Dataset:** [LIBERO-Mem](https://github.com/libero-mem/libero-mem)
- **Conversion Tool:** [any4lerobot](https://github.com/Tavish9/any4lerobot)
- **Conversion Process** [libero-mem](https://github.com/yyChen-sylvia/libero-mem)
- **Format:** LeRobot v2.1
- **Robot:** Panda Robot (7-DOF Arm + Gripper)
- **Tasks:** 10 long-term memory manipulation tasks.

## Dataset Structure
The dataset follows the standard LeRobot v2.1 structure:
- **`data/`**: Contains episode data in Parquet format.
- **`meta/`**: Contains task definitions, statistics, and camera configurations.
- ⚠️ **Note:**: `modality.json`, `stats_gr00t.json`, `steps_data_index.pkl` are intermediate products generated during the [StarVLA](https://github.com/starVLA/starVLA) training pipeline
- **`videos/`**: (If included) High-quality task execution videos.

### Key Observations/Actions
- `observation.images.agentview`: 256x256 agent-view camera RGB.
- `observation.images.eye_in_hand`: 256x256 wrist-mounted camera RGB.
- `observation.state`: Robot joint positions and gripper state.
- `action`: 7-DOF delta end-effector pose + gripper action.