Implementation:Isaac sim IsaacGymEnvs PBT Launcher

Knowledge Sources	IsaacGymEnvs
Domains	Experiment_Management, Distributed_Computing
Last Updated	2026-02-15 11:00 GMT

Overview

PBT_Launcher is the CLI entry point for the Population-Based Training (PBT) experiment launcher, routing experiment execution to local process, SLURM, or NGC backends based on user configuration.

Description

The PBT launcher module provides three key functions that together form the command-line interface for launching PBT experiments. The launcher_argparser() function constructs an ArgumentParser with common arguments including --train_dir (output directory), --run (Python module defining the experiment), --backend (execution backend choice), --pause_between (delay between process launches), and --experiment_suffix. It performs a partial parse to determine the selected backend, then dynamically adds backend-specific arguments by delegating to add_slurm_args(), add_ngc_args(), or add_os_parallelism_args().

The main() function orchestrates the full launch workflow. It parses command-line arguments, dynamically imports the user-specified run module (which must define a RUN_DESCRIPTION global variable of type RunDescription), applies the experiment suffix, and dispatches to the appropriate backend runner: run() for local OS processes, run_slurm() for SLURM cluster submission, or run_ngc() for NVIDIA GPU Cloud job submission.

The module serves as the unified entry point, abstracting away the differences between local multi-process execution, SLURM-based cluster scheduling, and NGC cloud deployment. This allows researchers to define experiments once and run them across different compute environments without changing the experiment description.

Usage

Use the PBT launcher as the CLI entry point for running hyperparameter sweeps or population-based training experiments. Invoke it directly with python -m isaacgymenvs.pbt.launcher.run --run your_experiment_module --backend processes. The experiment module must define a RUN_DESCRIPTION variable that specifies the experiments, parameter grids, and base commands.

Code Reference

Source Location

• Repository: IsaacGymEnvs
• File: isaacgymenvs/pbt/launcher/run.py
• Lines: 1-74

Signature

def launcher_argparser(args) -> argparse.ArgumentParser:
def parse_args():
def main():

Import

from isaacgymenvs.pbt.launcher.run import main

I/O Contract

Inputs

Name	Type	Required	Description
--run	str	Yes	Fully qualified Python module name that defines RUN_DESCRIPTION
--backend	str	No	Execution backend: processes (default), slurm, or ngc
--train_dir	str	No	Root directory for experiment outputs (default: ./train_dir)
--pause_between	int	No	Seconds to pause between launching processes (default: 1)
--experiment_suffix	str	No	Suffix to append to experiment directory names (default: empty)

Outputs

Name	Type	Description
return code	int	0 on success, 1 if the run module could not be imported

Usage Examples

# Command-line usage with local process backend:
# python -m isaacgymenvs.pbt.launcher.run \
#     --run isaacgymenvs.pbt.experiments.my_experiment \
#     --backend processes \
#     --train_dir ./train_dir \
#     --num_gpus 4 \
#     --max_parallel 8

# Command-line usage with SLURM backend:
# python -m isaacgymenvs.pbt.launcher.run \
#     --run isaacgymenvs.pbt.experiments.my_experiment \
#     --backend slurm \
#     --slurm_workdir ./slurm_workdir \
#     --slurm_gpus_per_job 1

# Programmatic usage:
from isaacgymenvs.pbt.launcher.run import main
exit_code = main()

Related Pages

• Isaac_sim_IsaacGymEnvs_RunDescription - Defines the experiment configurations consumed by the launcher
• Isaac_sim_IsaacGymEnvs_PBT_Process_Backend - Local OS process execution backend
• Isaac_sim_IsaacGymEnvs_PBT_Slurm_Backend - SLURM cluster execution backend
• Isaac_sim_IsaacGymEnvs_PBT_NGC_Backend - NGC cloud execution backend