Implementation:Isaac sim IsaacGymEnvs Hydra Task Train YAML

Knowledge Sources	IsaacGymEnvs Isaac Gym Docs
Type	Pattern Doc
Domains	Configuration, Training
Last Updated	2026-02-15 00:00 GMT

Overview

Concrete YAML configuration file templates for defining task environment parameters and training hyperparameters, with Cartpole as the reference example.

Description

Each new IsaacGymEnvs task requires two Hydra YAML files: a task config that defines the environment structure and physics settings, and a train config that defines the rl_games PPO training hyperparameters. This document provides complete templates for both files, annotated with explanations of each field.

Usage

Copy the templates below, adjust the parameters to match your task design specification, and place them in the correct directories: cfg/task/ for the task config and cfg/train/ for the training config.

Code Reference

Source Location

Repository: NVIDIA-Omniverse/IsaacGymEnvs
Task config: cfg/task/Cartpole.yaml (L1-45)
Train config: cfg/train/CartpolePPO.yaml (L1-68)
Top-level: cfg/config.yaml (L1-73)

Task Config Template

File: cfg/task/MyTask.yaml

# Task configuration for MyTask
# See IsaacGymEnvs cfg/task/Cartpole.yaml for reference

name: MyTask

physics_engine: ${..physics_engine}  # Inherited from top-level config

env:
  numEnvs: ${resolve_default:512,${...num_envs}}
  numObservations: 4       # Must match task's observation vector dimension
  numActions: 1            # Must match task's action vector dimension
  envSpacing: 4.0          # Spacing between parallel environments (meters)
  episodeLength: 500       # Maximum steps per episode
  enableDebugVis: False    # Enable debug visualization markers

  # Task-specific parameters
  resetDist: 3.0           # Distance threshold for cart reset
  maxEffort: 400.0         # Maximum force applied by actions

  # Clipping ranges (applied to observations before passing to policy)
  clipObservations: 5.0
  clipActions: 1.0

sim:
  dt: 0.0166               # Physics timestep (1/60 second)
  substeps: 2              # Physics substeps per control step
  up_axis: "z"             # Up axis (z for most tasks)
  use_gpu_pipeline: ${eq:${...pipeline},"gpu"}

  gravity:
    - 0.0
    - 0.0
    - -9.81

  physx:
    num_threads: ${....num_threads}
    solver_type: 1          # 0=PGS, 1=TGS
    num_position_iterations: 4
    num_velocity_iterations: 0
    contact_offset: 0.02    # Distance at which contacts are detected
    rest_offset: 0.001      # Distance at which contacts are resolved
    bounce_threshold_velocity: 0.2
    max_depenetration_velocity: 100.0
    default_buffer_size_multiplier: 2.0

  flex:
    num_outer_iterations: 4
    num_inner_iterations: 10
    warm_start: 0.25
    relaxation: 0.75

Cartpole Task Config Reference

File: cfg/task/Cartpole.yaml

# Reference: actual Cartpole task configuration
name: Cartpole

physics_engine: ${..physics_engine}

env:
  numEnvs: ${resolve_default:512,${...num_envs}}
  numObservations: 4
  numActions: 1
  envSpacing: 4.0
  episodeLength: 500
  enableDebugVis: False

  resetDist: 3.0
  maxEffort: 400.0

  clipObservations: 5.0
  clipActions: 1.0

sim:
  dt: 0.0166
  substeps: 2
  up_axis: "z"
  use_gpu_pipeline: ${eq:${...pipeline},"gpu"}
  gravity:
    - 0.0
    - 0.0
    - -9.81
  physx:
    num_threads: ${....num_threads}
    solver_type: 1
    num_position_iterations: 4
    num_velocity_iterations: 0
    contact_offset: 0.02
    rest_offset: 0.001
    bounce_threshold_velocity: 0.2
    max_depenetration_velocity: 100.0
    default_buffer_size_multiplier: 2.0

Train Config Template

File: cfg/train/MyTaskPPO.yaml

# Training configuration for MyTask using PPO via rl_games
# See IsaacGymEnvs cfg/train/CartpolePPO.yaml for reference

params:
  seed: ${...seed}

  algo:
    name: a2c_continuous       # rl_games algorithm name for continuous PPO

  model:
    name: continuous_a2c_logstd  # Policy model with learned log-std

  network:
    name: actor_critic           # Separate actor and critic networks
    separate: False              # False = shared backbone, True = separate networks

    space:
      continuous:
        mu_activation: None       # Output activation for mean (None = linear)
        sigma_activation: None
        mu_init:
          name: default
        sigma_init:
          name: const_initializer
          val: 0
        fixed_sigma: True

    mlp:
      units: [32, 32]            # Hidden layer sizes (small for Cartpole)
      activation: elu            # Activation function
      initializer:
        name: default
      regularizer:
        name: None

  config:
    name: ${resolve_default:MyTask,${....experiment}}
    full_experiment_name: ${.name}
    env_name: rlgpu             # Must be "rlgpu" for IsaacGymEnvs
    multi_gpu: ${....multi_gpu}
    ppo: True
    mixed_precision: False
    normalize_input: True        # Normalize observations
    normalize_value: True        # Normalize value function targets
    value_bootstrap: True
    num_actors: ${....task.env.numEnvs}
    reward_shaper:
      scale_value: 0.01         # Reward scaling factor

    # PPO hyperparameters
    gamma: 0.99                  # Discount factor
    tau: 0.95                    # GAE lambda
    learning_rate: 3e-4          # Adam learning rate
    lr_schedule: adaptive        # adaptive or linear
    kl_threshold: 0.008          # KL threshold for adaptive LR
    score_to_win: 20000          # Target score for early stopping
    max_epochs: ${resolve_default:500,${....max_iterations}}

    # Batch sizes
    horizon_length: 16           # Steps per environment per update
    minibatch_size: 8192         # Minibatch size for PPO updates
    mini_epochs: 8               # PPO epochs per update

    # PPO clip and regularization
    clip_param: 0.2              # PPO clip range
    entropy_coef: 0.0            # Entropy bonus coefficient
    e_clip: 0.2                  # Value function clip range
    truncate_grads: True
    grad_norm: 1.0               # Gradient norm clipping

Cartpole Train Config Reference

File: cfg/train/CartpolePPO.yaml

params:
  seed: ${...seed}

  algo:
    name: a2c_continuous

  model:
    name: continuous_a2c_logstd

  network:
    name: actor_critic
    separate: False
    space:
      continuous:
        mu_activation: None
        sigma_activation: None
        mu_init:
          name: default
        sigma_init:
          name: const_initializer
          val: 0
        fixed_sigma: True
    mlp:
      units: [32, 32]
      activation: elu
      initializer:
        name: default
      regularizer:
        name: None

  config:
    name: ${resolve_default:Cartpole,${....experiment}}
    full_experiment_name: ${.name}
    env_name: rlgpu
    ppo: True
    mixed_precision: False
    normalize_input: True
    normalize_value: True
    value_bootstrap: True
    num_actors: ${....task.env.numEnvs}
    reward_shaper:
      scale_value: 0.01
    gamma: 0.99
    tau: 0.95
    learning_rate: 3e-4
    lr_schedule: adaptive
    kl_threshold: 0.008
    score_to_win: 20000
    max_epochs: ${resolve_default:500,${....max_iterations}}
    horizon_length: 16
    minibatch_size: 8192
    mini_epochs: 8
    clip_param: 0.2
    entropy_coef: 0.0
    e_clip: 0.2
    truncate_grads: True
    grad_norm: 1.0

I/O Contract

Inputs

Name	Type	Required	Description
Task design spec	Document	Yes	Observation/action dimensions, episode length, reward parameters from design phase
Training hyperparameters	Values	Yes	Learning rate, network size, batch sizes, PPO settings

Outputs

Name	Type	Description
Task YAML	cfg/task/MyTask.yaml	Environment parameters composed into full config by Hydra
Train YAML	cfg/train/MyTaskPPO.yaml	Training hyperparameters composed into full config by Hydra

Naming Convention

File	Naming Pattern	Example
Task config	`cfg/task/{TaskName}.yaml`	`cfg/task/Cartpole.yaml`
Train config	`cfg/train/{TaskName}PPO.yaml`	`cfg/train/CartpolePPO.yaml`
Top-level default	`train: ${task}PPO`	Automatically resolves `CartpolePPO` from `task=Cartpole`

Related Pages

Isaac_sim_IsaacGymEnvs_Task_Configuration_Files - implements - Principle defining the separation of task and training configuration.
Isaac_sim_IsaacGymEnvs_Task_Design_Specification - prerequisite - Design specification providing the values for task config.
Isaac_sim_IsaacGymEnvs_Isaacgym_Task_Map_Registration - next step - After creating configs, register the task class.
Isaac_sim_IsaacGymEnvs_Train_Py_Task_Execution - consumer - Training pipeline that reads these configs.

Principle:Isaac_sim_IsaacGymEnvs_Task_Configuration_Files

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment