Implementation:Isaac sim IsaacGymEnvs Isaacgymenvs Make

Sources	Domains	Last Updated
IsaacGymEnvs	Environment, Simulation	2026-02-15 00:00 GMT

Overview

The top-level factory function that creates a fully initialized, GPU-accelerated vectorized RL environment from a task name string, handling configuration composition, simulation setup, and tensor buffer allocation.

Description

isaacgymenvs.make() is the public API for programmatic environment creation. It accepts a task name, device configuration, and optional Hydra config, then resolves the task class from the isaacgym_task_map registry, composes default configuration if none is provided, and instantiates the environment with the Isaac Gym simulation fully initialized and GPU tensor buffers allocated. The returned object is a VecTask subclass that exposes standard step(), reset(), and observe() methods.

Usage

Call from Python code when creating environments outside of the standard train.py workflow -- for example, in custom training loops, evaluation scripts, or Jupyter notebooks.

Code Reference

Source Location: Repository: NVIDIA-Omniverse/IsaacGymEnvs, File: isaacgymenvs/__init__.py (L14-55)

Signature:

def make(
    seed: int,
    task: str,
    num_envs: int,
    sim_device: str,
    rl_device: str,
    graphics_device_id: int = -1,
    headless: bool = False,
    multi_gpu: bool = False,
    virtual_screen_capture: bool = False,
    force_render: bool = True,
    cfg: DictConfig = None,
):
    """
    Creates a vectorized Isaac Gym environment.

    If cfg is None, composes default Hydra configuration for the
    specified task. Looks up the task class in isaacgym_task_map,
    instantiates it with the resolved config, and returns the
    fully initialized environment with GPU tensors acquired.
    """

Import:

import isaacgymenvs

I/O Contract

Inputs:

Parameter	Type	Required	Description
seed	int	Yes	Random seed for reproducibility
task	str	Yes	Task name string (must exist in isaacgym_task_map, e.g., "Cartpole", "Ant", "ShadowHand")
num_envs	int	Yes	Number of parallel environment instances to create on the GPU
sim_device	str	Yes	Device for physics simulation (e.g., "cuda:0", "cpu")
rl_device	str	Yes	Device for RL tensors -- observations, actions, rewards (e.g., "cuda:0")
graphics_device_id	int	No	GPU ID for rendering; -1 disables rendering (default -1)
headless	bool	No	If True, disables viewer window (default False)
multi_gpu	bool	No	Enable multi-GPU distributed training (default False)
virtual_screen_capture	bool	No	Enable virtual screen capture for headless rendering (default False)
force_render	bool	No	Force rendering even in headless mode for camera sensors (default True)
cfg	DictConfig	No	Pre-composed Hydra config; if None, default config is composed automatically

Outputs:

Instantiated VecTask subclass with:
- Isaac Gym simulation (sim) created and configured
- num_envs parallel environments populated with loaded assets and actors
- GPU tensor buffers allocated for observations, actions, rewards, and dones
- Standard environment API methods available: step(actions), reset(), observe()

Usage Examples

Example 1 -- Create a Cartpole environment programmatically:

import isaacgymenvs
import torch

envs = isaacgymenvs.make(
    seed=42,
    task="Cartpole",
    num_envs=256,
    sim_device="cuda:0",
    rl_device="cuda:0",
    graphics_device_id=0,
    headless=True,
)

print(f"Observation space: {envs.observation_space}")
print(f"Action space: {envs.action_space}")

# Run a random policy
obs = envs.reset()
for step in range(1000):
    actions = torch.randn(256, envs.action_space.shape[0], device="cuda:0")
    obs, rewards, dones, info = envs.step(actions)

Example 2 -- Create environment with custom config:

import isaacgymenvs
from omegaconf import OmegaConf

# Load and modify config
cfg = OmegaConf.load("cfg/config.yaml")
cfg.task_name = "Ant"
cfg.num_envs = 4096

envs = isaacgymenvs.make(
    seed=0,
    task="Ant",
    num_envs=4096,
    sim_device="cuda:0",
    rl_device="cuda:0",
    headless=True,
    cfg=cfg,
)

Example 3 -- Headless environment for evaluation:

import isaacgymenvs

envs = isaacgymenvs.make(
    seed=123,
    task="ShadowHand",
    num_envs=64,
    sim_device="cuda:0",
    rl_device="cuda:0",
    graphics_device_id=-1,
    headless=True,
    force_render=False,
)

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment