Implementation:Isaac sim IsaacGymEnvs Checkpoint Filesystem Convention

Sources	Domains	Last Updated
IsaacGymEnvs, common_agent.py	Persistence, Training	2026-02-15 00:00 GMT

Overview

Pattern for saving and discovering trained RL policy checkpoints following the rl_games filesystem convention within IsaacGymEnvs.

Description

The CommonAgent class in its train() method handles periodic checkpoint saving. Checkpoints are written as .pth files containing the model state_dict to a predictable directory path derived from the experiment name configured via Hydra.

Usage

Checkpoints are produced automatically during training runs and consumed when resuming training or launching inference. No explicit path configuration is required beyond specifying the experiment name.

Code Reference

Source Location: Repository: NVIDIA-Omniverse/IsaacGymEnvs, File: isaacgymenvs/learning/common_agent.py (L122-176)

Pattern:

runs/<experiment_name>/nn/<checkpoint>.pth

Checkpoint Saving Code (common_agent.py train() method):

# In CommonAgent.train():
# Periodic checkpoint saving based on save_freq
if self.epoch_num % self.save_freq == 0:
    self.save(os.path.join(self.nn_dir,
        'last_' + self.config['name'] + 'ep' + str(self.epoch_num)
        + 'rew' + str(mean_rewards)))

# Best model checkpoint saving
if mean_rewards > self.last_mean_rewards:
    self.last_mean_rewards = mean_rewards
    self.save(os.path.join(self.nn_dir, self.config['name']))

I/O Contract

Inputs:

Input	Type	Description
Completed training epoch	Internal state	Triggers checkpoint save evaluation
save_freq	int	How often (in epochs) to save periodic checkpoints
experiment_name	str	Derived from Hydra config; determines directory naming

Outputs:

Output	Type	Description
Best checkpoint	.pth file	runs/<experiment_name>/nn/<experiment_name>.pth -- saved when mean reward improves
Periodic checkpoint	.pth file	runs/<experiment_name>/nn/last_<name>ep<N>rew<R>.pth -- saved every save_freq epochs

Directory Layout Example

runs/
  Cartpole/
    nn/
      Cartpole.pth                     # best checkpoint
      last_Cartpoleep100rew485.3.pth   # periodic checkpoint at epoch 100
      last_Cartpoleep200rew499.1.pth   # periodic checkpoint at epoch 200
    summaries/
      events.out.tfevents.*            # TensorBoard logs

Related Pages

• Principle:Isaac_sim_IsaacGymEnvs_Checkpoint_Location