README.md

---
dataset_name: LIBERO-Mem
pretty_name: "LIBERO-Mem: Long-Horizon Object-Centric Kitchen Manipulation"
tags:
  - robotics
  - imitation-learning
  - reinforcement-learning
  - manipulation
  - vision
  - hdf5
  - datasets
license: mit
task_categories:
  - reinforcement-learning
  - other
language:
  - en
---

# LIBERO-Mem Dataset Specification  
### Metadata (`metainfo.json`) + Demonstrations (`.hdf5`)

This document presents the complete schema for the **LIBERO-Mem dataset**, covering both:

1. **`metainfo.json`** — task-level metadata, bounding boxes, segmentation, and initial states  
2. **`.hdf5` demonstration files** — synchronized RGB-D observations, segmentation maps, proprioception, and control actions

Both sources together provide a **time-aligned**, **object-centric**, and **pixel-level** representation of robot manipulation trajectories.

---

# 📘 Part I — `metainfo.json` Metadata Format

The `metainfo.json` file contains **all task-level metadata** used to interpret and reconstruct demonstrations.  
Each top-level key represents a task, and inside each task are one or more demonstration entries (`demo_1`, `demo_2`, …).

---

## 🌟 Top-Level Structure

```json
{
  "<TASK_NAME>": {
    "demo_1": { ... },
    "demo_2": { ... }
  },
  "<TASK_NAME_2>": { ... }
}
```

Examples of task names:

- `KITCHEN_SCENE1_1_pick_up_the_bowl_and_place_it_back_on_the_plate`
- `KITCHEN_SCENE1_7_swap_the_2_bowls_on_their_plates_using_the_empty_plate`
- `KITCHEN_SCENE1_9_put_the_cream_cheese_in_the_nearest_basket_and_place_that_basket_in_the_center`

---

## 📁 Per-Demo Structure

Each demonstration contains six fields:

### 1. **success**
- **Type:** `bool`  
Indicates whether the demonstration completes its intended task.

### 2. **initial_state**
- **Type:** `list[number]`  
Simulator state vector for restoring the initial conditions.

### 3. **task_nouns**
- **Type:** `list[string]`  
Core object references for the task.

### 4. **task_description**
- **Type:** `string`  
Natural-language description of the task.

### 5. **exo_boxes**
- **Type:** `list[frame_dict]`  
Bounding boxes from the **exo-camera**, one per timestep.

### 6. **ego_boxes**
- **Type:** `list[frame_dict]`  
Bounding boxes from the **ego-camera**, same structure as `exo_boxes`.

---

## 🧩 Metadata Summary

```
<TASK>/<DEMO>/
    success: bool
    initial_state: number[N]
    task_nouns: string[K]
    task_description: string
    exo_boxes: list<frame_dict>
    ego_boxes: list<frame_dict>

frame_dict:
    "<object>": [seg_id, [cx, cy, w, h], obj_subgoal]
```

---

# 📘 Part II — HDF5 Demonstration Format

Each `.hdf5` file stores raw observations, proprioception, and control data.

---

## 📁 File Structure

```
data/
    demo_0/
    demo_1/
    demo_2/
```

---

## 🔧 Per-Demo Structure

```
data/demo_i/
    actions               (T, 7)
    dones                 (T,)
    obs/
        agentview_rgb
        agentview_depth
        agentview_seg
        eye_in_hand_rgb
        eye_in_hand_depth
        eye_in_hand_seg
        gripper_states
        joint_states
```

---

## 📊 Field Descriptions

### actions  
`(T, 7)` — float64

### dones  
`(T,)` — uint8

### Observation Fields
- agentview_rgb — `(T,256,256,3)`
- agentview_depth — `(T,256,256)`
- agentview_seg — `(T,256,256)`
- eye_in_hand_rgb — `(T,256,256,3)`
- eye_in_hand_depth — `(T,256,256)`
- eye_in_hand_seg — `(T,256,256)`

### Proprioception
- gripper_states — `(T,2)`
- joint_states — `(T,7)`

---

# 📘 Combined Schema Overview

```
metainfo.json
    <TASK>/<DEMO>/
        success: bool
        initial_state: number[N]
        task_nouns: string[3]
        task_description: string
        exo_boxes: list<frame_dict>
        ego_boxes: list<frame_dict>

HDF5
    data/demo_i/
        actions: (T, 7)
        dones: (T,)
        obs/
            agentview_rgb
            agentview_depth
            agentview_seg
            eye_in_hand_rgb
            eye_in_hand_depth
            eye_in_hand_seg
            gripper_states
            joint_states
```