Principle:ARISE Initiative Robomimic Configuration Generation
| Knowledge Sources | |
|---|---|
| Domains | Robotics, Experiment_Management, Hyperparameter_Optimization |
| Last Updated | 2026-02-15 08:00 GMT |
Overview
A combinatorial configuration generation pattern that computes the Cartesian product of grouped parameter ranges, writes individual JSON configs, and generates a launch shell script for parallel training.
Description
Configuration Generation is the execution phase of hyperparameter sweeping. After parameters are defined (via Sweep_Parameter_Definition), this principle computes all valid parameter combinations, writes one JSON config file per combination, and generates a bash script that launches a training run for each config.
The generation process follows three stages:
- Range computation: Group parameters by group ID, compute Cartesian product across groups while zipping within groups
- JSON generation: For each combination, clone the base config, apply parameter overrides, set the experiment name, and write to a JSON file
- Script generation: Write a bash script with one training command per config file
The experiment name for each config is automatically constructed from the parameter names and values, making it easy to identify configurations from directory names.
Usage
Use this principle after defining sweep parameters via ConfigGenerator.add_param. Call generate() to produce all config files and the launch script, then execute the script (possibly with GPU parallelization) to run the sweep.
Theoretical Basis
# Abstract generation pipeline (not real implementation)
# Stage 1: Compute combinations
groups = group_parameters_by_id(parameters)
ranges = cartesian_product_across_groups(groups)
# If group 1 has [a,b] and group 2 has [x,y,z]:
# ranges = [(a,x), (a,y), (a,z), (b,x), (b,y), (b,z)]
# Stage 2: Generate JSON configs
for combination in ranges:
config = load_base_config()
for param, value in combination:
config[param.key] = value
config["experiment"]["name"] = build_experiment_name(combination)
write_json(config, path)
# Stage 3: Generate launch script
write_script("python train.py --config <path>" for each config)