dancher00/maniskill-panda-pickcube

ManiSkill Panda PickCube Dataset

This dataset contains robot demonstrations for pick-and-place tasks using a Franka Panda robot in the ManiSkill simulation environment.

Dataset Description

This dataset was collected using ManiSkill PickCube-v1 environment and converted to LeRobot format for training Vision-Language-Action (VLA) models, specifically optimized for pi0 architecture.

Task Description

The robot needs to pick up a cube and place it in a designated location. The task involves:

• Visual perception of the cube and target location
• Precise manipulation and grasping
• Coordinated arm and gripper movements

Robot Specifications

• Robot: Franka Panda (7-DOF manipulator)
• End-effector: Parallel gripper
• Control: Cartesian space control with gripper commands
• Observation: RGB camera + joint states

Dataset Statistics

• Total Episodes: 100
• Total Frames: 5,000
• Average Episode Length: 50.0 frames
• Frame Rate: 30 FPS
• Action Dimension: 29D (x, y, z, rx, ry, rz, gripper)
• State Dimension: 7D (joint positions)
• Image Resolution: 224×224

Data Format

The dataset follows the LeRobot standard format:

Features

• action: Robot actions in Cartesian space + gripper command

  • Shape: [29]
  • Format: [x, y, z, rx, ry, rz, gripper]
  • Units: positions in meters, orientations in radians, gripper in [0,1]

• observation.state: Joint positions of the robot

  • Shape: [7]
  • Format: [joint_1, joint_2, ..., joint_7]
  • Units: radians

• observation.images.main: RGB camera observations

  • Shape: [224, 224, 3]
  • Format: RGB images
  • Encoding: MP4 videos

Metadata

• timestamp: Time elapsed since episode start (seconds)
• frame_index: Frame number within episode
• episode_index: Episode identifier
• index: Global frame index across all episodes
• task_index: Task type identifier (always 0 for this dataset)

Usage

Loading with LeRobot

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

# Load the dataset
dataset = LeRobotDataset("dancher00/maniskill-panda-pickcube")

# Access data
for episode_idx in range(len(dataset.episode_data_index["episode_index"])):
    episode = dataset[episode_idx]
    actions = episode["action"]
    observations = episode["observation.state"] 
    images = episode["observation.images.main"]

Training with pi0

from lerobot.common.policies.pi0.modeling_pi0 import Pi0Policy

# Initialize policy
policy = Pi0Policy(
    config={
        "action_dim": 29,
        "state_dim": 7,
        "chunk_size": 50,
        "n_action_steps": 50,
        "resize_imgs_with_padding": [224, 224],
        # ... other pi0 config parameters
    }
)

# Train on dataset
trainer.train(policy, dataset)

Dataset Structure

dancher00/maniskill-panda-pickcube/
├── data/
│   └── chunk-000/
│       ├── episode_000000.parquet
│       ├── episode_000001.parquet
│       └── ...
├── videos/
│   └── chunk-000/
│       └── observation.images.main/
│           ├── episode_000000.mp4
│           ├── episode_000001.mp4
│           └── ...
├── meta/
│   ├── episodes.jsonl
│   ├── info.json
│   ├── stats.json
│   └── tasks.jsonl
└── README.md

Technical Details

Data Collection

• Simulator: ManiSkill 2.0
• Environment: PickCube-v1
• Collection Method: Demonstrations
• Success Rate: Varies by episode

Data Processing

• Image Preprocessing: Resized to 224×224 with aspect ratio preservation
• Action Space: Continuous Cartesian control
• State Representation: Joint positions only
• Temporal: 30 FPS sampling rate

Quality Assurance

• All episodes validated for completeness
• Action and state bounds checked
• Video integrity verified
• Statistics computed and validated

Intended Use

This dataset is designed for:

• Training VLA models (especially pi0 architecture)
• Robotics research in manipulation
• Benchmarking pick-and-place algorithms
• Transfer learning to real robot systems

Limitations

• Simulation only: Collected in ManiSkill, may need domain adaptation for real robots
• Single task: Only pick-and-place variations
• Limited diversity: Fixed robot and gripper setup
• Scale: Relatively small dataset (100 episodes)

License

This dataset is released under the Apache 2.0 license.

Citation

If you use this dataset in your research, please cite:

@dataset{maniskill_panda_pickcube_2025,
  title={ManiSkill Panda PickCube Dataset},
  author={Anonymous},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/dancher00/maniskill-panda-pickcube}
}

Acknowledgments

• ManiSkill: For providing the simulation environment
• LeRobot: For the dataset format and training framework
• Hugging Face: For hosting and distribution infrastructure