Principle:Isaac sim IsaacGymEnvs Inference Configuration
| Sources | Domains | Last Updated |
|---|---|---|
| IsaacGymEnvs, train.py | Configuration, Inference | 2026-02-15 00:00 GMT |
Overview
Configuration pattern for switching the training pipeline to inference mode by setting test flags and specifying checkpoint paths.
Description
The same train.py entry point serves both training and inference. By setting test=True and providing a checkpoint path, the system creates a Player instead of an Agent. Additional parameters like sigma can override exploration noise. The config composition is identical to training but routes to the play path.
This design means users do not need separate scripts or configurations for evaluation -- the full Hydra config infrastructure (task selection, environment parameters, network architecture) is reused. The only difference is the final dispatch: train=False, play=True instead of train=True, play=False.
Key configuration parameters for inference mode:
| Parameter | Default | Description |
|---|---|---|
| test | False | When True, switches to inference/evaluation mode |
| checkpoint | Path to the .pth checkpoint file to load | |
| sigma | None | Optional override for exploration noise (set to 0.0 for deterministic) |
| num_envs | (from task config) | Can be reduced for visualization or increased for batch evaluation |
Usage
When evaluating a trained policy or visualizing its behavior. Typical invocation:
python train.py test=True checkpoint=runs/Cartpole/nn/Cartpole.pth task=Cartpole num_envs=64
Theoretical Basis
Single entry point with mode flag -- reuse config composition infrastructure for both training and evaluation. This avoids config duplication and ensures that the same environment parameters, network architecture, and normalization settings used during training are faithfully reproduced during inference.