Link Search Menu Expand Document
gym2real

Gym2Real_isaacgym

https://github.com/kevinh42/gym2real_isaacgym

This repo provides a custom task to be used with Isaac Gym as well as scripts for training and exporting models.

Repo Structure

  • train.py: Main script used for training/testing models
  • export.py: Script for exporting models as ONNX files
  • cfg/task/: YAML configs for each task defining environment parameters and domain randomization
  • cfg/train/: YAML configs for each task defining RL model and hyperparameters
  • tasks/: Class definitions and assets for each task
  • runs/: Model checkpoints and summaries will be saved here

Usage

Adding a New Task

Create a new folder called tasks/{task_name}. The following files will be needed:

  • tasks/{task_name}/{task_name}.py: Class defining the task. Rewards, observations, actions are all defined here.
  • tasks/{task_name}/assets/: URDF and meshes should be placed here.
  • cfg/task/{TaskName}.yaml: Config file defining environment parameters and domain randomization
  • cfg/train/{TaskName}PPO.yaml: Config gile defining hyperparameters for training

Training

  • python train.py task={task_name} to begin training
  • Models are saved as runs/{TaskName}/nn/{checkpoint_name}.pth

Exporting ONNX

  • python export.py task={task_name} checkpoint="runs/{TaskName}/nn/{checkpoint_name}.pth" to export
  • Model is exported to runs/{TaskName}/nn/{checkpoint_name}.pth.onnx

train.py

__init__

Loads configured parameters and sets up tensors.

_create_envs

Loads assets and creates agents. DOFs (joints) and rigid bodies (links) are configured here.

compute_reward

Calculates reward from observations.

compute_observations

Data from environments are placed in the observations buffer for calculating reward and updating neural network.

reset_idx

Anything that should be applied to the agent on reset can be added here (eg. randomization, initial pose, initial DOF states)

pre_physics_step

This step occurs after the network has been updated and has outputted an action. This action can be applied to control the agent here (eg. position/velocity/effort control of DOFs). Physics simulation happens after this step.

post_physics_step

This step occurs after any physics simulation has finished. Observations and rewards are computed here for the next step.