Hierarchical Data Format version 5

A general-purpose file format that stores many large arrays hierarchically in one file, commonly used for robot demonstration data.

HDF5 is an open-source file format and software library maintained by the HDF Group, usually with a .h5 or .hdf5 extension. It works like a folder system inside a single file: “groups” provide the hierarchy, each group holds multidimensional arrays (called datasets), metadata can be attached anywhere, and any part of the file can be read without loading the rest. A robot demonstration contains multiple camera streams, joint state, and actions as time series, which fits naturally into HDF5. robomimic stores a whole dataset this way, with each trajectory as a group like data/demo_0; the ACT/ALOHA collection scripts likewise save each episode as its own HDF5 file. Storing images verbatim makes files large, so newer open datasets increasingly use the LeRobot format or RLDS instead.

ExampleIn a robomimic dataset file, data/demo_0/actions is one trajectory's action array, shaped (number of steps, action dimension); data/demo_0/obs/agentview_image holds the camera image for each of those same time steps.

Also called

HDF5, .h5 File, .hdf5 File

Related

LeRobotDataset · RLDS (Reinforcement Learning Datasets) · Zarr · Apache Parquet · robomimic · Action Chunking with Transformers

Sources

The HDF Group: HDF5
robomimic Docs: Datasets Overview
ACT (tonyzhaozh/act) GitHub