README.md

metadata

license: apache-2.0
task_categories:
  - robotics
tags:
  - robotics
  - manipulation
  - libero
  - libero-pro
  - imitation-learning
  - vla
size_categories:
  - 1K<n<10K

Robotics Dataset

Synthetic demonstration dataset for LIBERO-Pro robot manipulation, generated with the OpenRobot massive dataset pipeline.

Contents

• 4,800 episodes across 16 LIBERO-Pro suites (300 episodes each)
• ~77 GB of compressed .npz demonstrations
• 30 episodes per task × 10 tasks per suite

Suites

Suite	Episodes
libero_spatial	300
libero_spatial_lan	300
libero_spatial_object	300
libero_spatial_swap	300
libero_object	300
libero_object_lan	300
libero_object_object	300
libero_object_swap	300
libero_goal	300
libero_goal_lan	300
libero_goal_object	300
libero_goal_swap	300
libero_10	300
libero_10_lan	300
libero_10_object	300
libero_10_swap	300

Episode format (.npz)

Each file is a compressed NumPy archive with:

Key	Shape	Description
image	(T, 224, 224, 3)	Third-person RGB observations
wrist_image	(T, 224, 224, 3)	Wrist camera RGB observations
state	(T, 8)	Proprioceptive state
actions	(T, 7)	Robot actions (6-DOF + gripper)
task	scalar str	Natural-language task instruction
suite	scalar str	LIBERO-Pro suite name

Files are named ep_t{task_id:02d}_{trial:04d}.npz inside per-suite subdirectories.

Generation

Produced by generate_massive_dataset.py using privileged MuJoCo oracle trajectories, swap/standard layouts, linguistic paraphrases, and initial-state randomization.

Usage

import numpy as np

ep = np.load("libero_spatial/ep_t00_0010.npz", allow_pickle=True)
images = ep["image"]       # (T, 224, 224, 3)
actions = ep["actions"]    # (T, 7)
instruction = str(ep["task"])