Implementation:ARISE Initiative Robomimic ConfigGenerator generate

Knowledge Sources	robomimic robomimic Hyperparam Scan
Domains	Robotics, Experiment_Management, Hyperparameter_Optimization
Last Updated	2026-02-15 08:00 GMT

Overview

Concrete tool for generating combinatorial JSON config files and a launch script from accumulated parameter sweep definitions provided by the robomimic hyperparameter utilities module.

Description

The ConfigGenerator.generate method orchestrates the full sweep generation pipeline. It calls _generate_jsons (L202-292) to compute parameter combinations and write JSON configs, then _script_from_jsons (L294-308) to produce the launch bash script.

Internally, _get_parameter_ranges (L128-200) groups parameters by group ID, computes Cartesian product across groups while zipping within groups, and returns all combinations. _generate_jsons loads the base config, applies each combination's overrides using hierarchical key parsing (slash-separated keys), sets experiment names, and writes JSON files.

Usage

Call this method after accumulating all parameters via add_param. Execute the resulting shell script to launch the training sweep.

Code Reference

Source Location

• Repository: robomimic
• File: robomimic/utils/hyperparam_utils.py
• Lines: L80-89 (generate), L128-200 (_get_parameter_ranges), L202-292 (_generate_jsons), L294-308 (_script_from_jsons)

Signature

def generate(self, override_base_name=False):
    """
    Generates json configs for the hyperparameter sweep using attributes
    @self.parameters, @self.base_config_file, and @self.script_file.

    Args:
        override_base_name (bool): if True, use empty string as base experiment name
    """

Import

from robomimic.utils.hyperparam_utils import ConfigGenerator

# Usage:
generator = ConfigGenerator(base_config_file="bc.json", script_file="sweep.sh")
generator.add_param(...)
generator.generate()

I/O Contract

Inputs

Name	Type	Required	Description
self.parameters	OrderedDict	Yes	Accumulated parameter definitions from add_param calls
self.base_config_file	str	Yes	Path to base JSON config
self.script_file	str	Yes	Output shell script path
override_base_name	bool	No	Use empty base name. Default: False

Outputs

Name	Type	Description
JSON config files	Files	One JSON file per parameter combination in generated_config_dir (or same dir as base config)
Shell script	File	Bash script at script_file with one "python train.py --config <path>" per config

Usage Examples

Complete Sweep Workflow

from robomimic.utils.hyperparam_utils import ConfigGenerator

# 1. Create generator
generator = ConfigGenerator(
    base_config_file="robomimic/exps/templates/bc.json",
    script_file="/tmp/run_sweep.sh",
    generated_config_dir="/tmp/sweep_configs/",
)

# 2. Define sweep parameters
generator.add_param(
    key="algo/optim_params/policy/learning_rate/initial",
    name="lr", group=1, values=[1e-3, 1e-4],
)
generator.add_param(
    key="train/num_epochs", name="ep", group=2, values=[100, 200],
)

# 3. Generate configs (2 × 2 = 4 configs)
generator.generate()

# 4. Launch sweep
# bash /tmp/run_sweep.sh

Launch Generated Script

# Run all sweep configs sequentially
bash /tmp/run_sweep.sh

# Or run in parallel across GPUs
cat /tmp/run_sweep.sh | xargs -I {} -P 4 bash -c '{}'

Related Pages

Implements Principle

• Principle:ARISE_Initiative_Robomimic_Configuration_Generation