Implementation:Isaac sim IsaacGymEnvs Launch Rlg Hydra Inference
| Sources | Domains | Last Updated |
|---|---|---|
| IsaacGymEnvs, train.py | Configuration, Inference | 2026-02-15 00:00 GMT |
Overview
The concrete entry point and dispatch logic within train.py that routes execution to inference (play) mode when test=True is specified.
Description
The same launch_rlg_hydra(cfg) function used for training also handles the inference path. When cfg.test is True, the function constructs a run configuration dict with train=False and play=True, passing the checkpoint path and sigma override to the Runner.
Usage
Invoked via CLI with test=True and a checkpoint path. The Hydra config composition resolves all task/train parameters identically to training mode.
Code Reference
Source Location: Repository: NVIDIA-Omniverse/IsaacGymEnvs, File: isaacgymenvs/train.py (L71-72 hydra entry, L202-215 test/play dispatch)
Signature:
def launch_rlg_hydra(cfg: DictConfig):
"""Main entry point - handles both training and inference."""
# ... config setup, environment creation, runner building ...
# Test/play dispatch (L202-215):
if cfg.test:
runner.run({
'train': False,
'play': True,
'checkpoint': cfg.checkpoint,
'sigma': cfg.sigma
})
else:
runner.run({
'train': True,
'play': False
})
CLI Invocation:
python train.py test=True checkpoint=<path> task=<TaskName> num_envs=<N>
I/O Contract
Inputs:
| Parameter | Type | Required | Description |
|---|---|---|---|
| cfg.test | bool | Yes | Must be True to trigger inference path |
| cfg.checkpoint | str | Yes | Path to the .pth checkpoint file |
| cfg.task | str | Yes | Task name (e.g., Cartpole, Ant, Humanoid) |
| cfg.sigma | float | No | Override for exploration noise; None uses model default |
| cfg.num_envs | int | No | Number of parallel environments for inference |
Outputs:
- Config dict that triggers play mode in Runner
- Runner dispatches to the registered Player class for the given algorithm
Usage Examples
Example 1 -- Basic inference with a trained Ant policy:
python train.py test=True checkpoint=runs/Ant/nn/Ant.pth task=Ant num_envs=64
Example 2 -- Deterministic inference (zero exploration noise):
python train.py test=True checkpoint=runs/Humanoid/nn/Humanoid.pth task=Humanoid sigma=0.0