RoboCerebra_hdf5 / README.md
CollisionCode's picture
Update README.md
89e635e verified
metadata
license: apache-2.0
task_categories:
  - robotics
tags:
  - libero
  - lerobot
  - embodied-ai
  - manipulation
pretty_name: RoboCerebra (HDF5 Format)
size_categories:
  - 1G<n<500G

πŸ“¦ RoboCerebra HDF5 Dataset

License Project Page

This repository contains the structured HDF5 dataset for the RoboCerebra project, optimized for high-performance robot learning training.

βœ… Status: Data Upload Complete. All large-scale data files are now available for full download and experimental use.

πŸ“– Data Source

The dataset is derived from the RoboCerebra-Bench raw collection.

  • Content: Humanoid robotic manipulation demonstrations across various complex environments and tasks.
  • Structure: Raw trajectory data has been processed and unified into a standard HDF5 format, enabling fast random access and efficient data loading during training.

πŸ› οΈ Data Conversion Script

The HDF5 files in this repository were converted from the official raw dataset qiukingballball/RoboCerebraBench using the processing scripts provided in the official RoboCerebra repository.

If you need to regenerate the HDF5 files or process custom data subsets, please follow the instructions from the official source:

  1. Clone the official conversion tool: Visit github.com/qiuboxiang/RoboCerebra to get the latest scripts.

  2. Navigate to the builder directory:

    cd rlds_dataset_builder/
    
  3. Run the conversion script: Ensure the --robocerebra_raw_data_dir points to your local copy of the RoboCerebraBench raw data.

    python regenerate_robocerebra_dataset.py \
      --robocerebra_raw_data_dir "/path/to/RoboCerebraBench/Ideal" \
      --robocerebra_target_dir "./converted_hdf5/robocerebra_ideal"
    

Argument Descriptions

  • --robocerebra_raw_data_dir: Root directory containing the raw demonstration data from qiukingballball/RoboCerebraBench.
  • --robocerebra_target_dir: Target path for the output HDF5 files.