Principle:Isaac sim IsaacGymEnvs Checkpoint Location
| Sources | Domains | Last Updated |
|---|---|---|
| IsaacGymEnvs, common_agent.py | Persistence, Training | 2026-02-15 00:00 GMT |
Overview
Convention for organizing and locating trained RL policy checkpoint files within the IsaacGymEnvs directory structure.
Description
Trained model checkpoints are saved as .pth files under runs/<experiment_name>/nn/ following rl_games conventions. The experiment name is derived from the Hydra config (typically the task name or a custom experiment field). Checkpoints contain the model state_dict and are saved periodically during training based on the save_freq parameter. Users locate these files to resume training or run inference.
The directory layout follows a predictable hierarchy:
runs/
<experiment_name>/
nn/
<experiment_name>.pth # latest/best checkpoint
last_<experiment_name>ep<N>rew<R>.pth # periodic checkpoints
summaries/ # TensorBoard logs
This convention enables checkpoint discovery without explicit path specification -- users only need to know the experiment name to locate any associated artifacts.
Usage
When loading a previously trained policy for evaluation, inference, or continued training. The checkpoint path is passed via the checkpoint CLI argument or resolved automatically by the Runner when resuming.
Theoretical Basis
Convention over configuration -- a standardized directory layout enables checkpoint discovery without explicit path specification. By following the rl_games convention, IsaacGymEnvs ensures compatibility with the upstream training framework's save/load mechanics. The periodic checkpoint naming scheme (embedding epoch number and reward) provides a human-readable audit trail of training progress.