Implementation:ARISE Initiative Robomimic Split train val from hdf5

Knowledge Sources	robomimic
Domains	Robotics, Data_Pipeline, Data_Splitting
Last Updated	2026-02-15 08:00 GMT

Overview

Concrete tool for splitting HDF5 demonstration datasets into training and validation subsets using filter keys provided by the robomimic split_train_val script.

Description

The split_train_val_from_hdf5 function creates randomized train/valid splits at the demonstration level. It reads available demonstrations (from the full dataset or a specified filter key subset), randomly assigns them to train and valid sets based on the specified ratio, and writes the resulting demo lists as filter keys in the HDF5 mask/ group via create_hdf5_filter_key.

Usage

Run as a CLI script after observation extraction. The resulting "train" and "valid" (or "{filter_key}_train" / "{filter_key}_valid") keys are used by the training pipeline's dataset loading step.

Code Reference

Source Location

• Repository: robomimic
• File: robomimic/scripts/split_train_val.py
• Lines: L25-76

Signature

def split_train_val_from_hdf5(hdf5_path, val_ratio=0.1, filter_key=None):
    """
    Splits data into training set and validation set from HDF5 file.

    Args:
        hdf5_path (str): path to the hdf5 file to load the transitions from
        val_ratio (float): ratio of validation demonstrations to all demonstrations
        filter_key (str): if provided, split the subset of demonstration keys stored
            under mask/@filter_key instead of the full set of demonstrations
    """

Import

from robomimic.scripts.split_train_val import split_train_val_from_hdf5

I/O Contract

Inputs

Name	Type	Required	Description
hdf5_path	str	Yes	Path to HDF5 dataset file
val_ratio	float	No	Fraction of demos for validation. Default: 0.1
filter_key	str	No	If provided, split only demos in mask/{filter_key}

Outputs

Name	Type	Description
(side effect)	None	Writes "train" and "valid" filter keys (or "{filter_key}_train" / "{filter_key}_valid") into HDF5 mask/ group

Usage Examples

CLI Usage

# Split with default 10% validation
python robomimic/scripts/split_train_val.py --dataset /path/to/low_dim_v141.hdf5

# Split with 20% validation
python robomimic/scripts/split_train_val.py --dataset /path/to/low_dim_v141.hdf5 --ratio 0.2

# Split a subset
python robomimic/scripts/split_train_val.py --dataset /path/to/low_dim_v141.hdf5 --filter_key 50_demos

Programmatic Usage

from robomimic.scripts.split_train_val import split_train_val_from_hdf5

split_train_val_from_hdf5(
    hdf5_path="/path/to/low_dim_v141.hdf5",
    val_ratio=0.1,
)
# Creates mask/train and mask/valid in the HDF5 file

Related Pages

Implements Principle

• Principle:ARISE_Initiative_Robomimic_Train_Validation_Split

Requires Environment

• Environment:ARISE_Initiative_Robomimic_HDF5_Data_Dependencies