README.md

metadata

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

{
  "<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