envpool

---

EnvPool is a C++-based batched environment pool with pybind11 and thread pool. It has high performance (~1M raw FPS with Atari games, ~3M raw FPS with MuJoCo simulator on DGX-A100) and compatible APIs (supports Gymnasium and dm_env, both sync and async, both single and multi player environment). Currently it supports:

• Atari games
• MuJoCo (Gymnasium)
• Classic control RL envs: CartPole, MountainCar, Pendulum, Acrobot
• Toy text RL envs: Catch, FrozenLake, Taxi, NChain, CliffWalking, Blackjack
• ViZDoom single player
• DeepMind Control Suite
• Box2D
• Google Research Football
• Procgen
• Minigrid
• MarlGrid
• Highway
• MetaWorld
• MyoSuite
• PGX
• Jumanji
• MuJoCo Playground

Here are EnvPool's several highlights:

• All-platform support: Linux, macOS, and Windows wheels for Python 3.11-3.14;
• Built-in rendering with batched rgb_array output and human display mode;
• Compatible with gymnasium APIs and DeepMind dm_env APIs;
• Manage a pool of envs, interact with the envs in batched APIs by default;
• Support both synchronous execution and asynchronous execution;
• Support both single player and multi-player environment;
• Easy C++ developer API to add new envs: Customized C++ environment integration;
• Free ~2x speedup with only single environment;
• 1 Million Atari frames / 3 Million MuJoCo steps per second simulation with 256 CPU cores, ~20x throughput of Python subprocess-based vector env;
• ~3x throughput of Python subprocess-based vector env on low resource setup like 12 CPU cores;
• Comparing with existing GPU-based solution (Brax / Isaac-gym), EnvPool is a general solution for various kinds of speeding-up RL environment parallelization;
• XLA support with JAX jit function: XLA Interface;
• Compatible with some existing RL libraries, e.g., Stable-Baselines3, Tianshou, ACME, CleanRL, or rl_games.
  • Stable-Baselines3 Pendulum-v1 example;
  • Tianshou CartPole example, Pendulum-v1 example, Atari example, MuJoCo example, and integration guideline;
  • ACME HalfCheetah example;
  • CleanRL Pong-v5 example (Solving Pong in 5 mins (tracked experiment));
  • rl_games Atari example (2 mins Pong and 15 mins Breakout) and MuJoCo example (5 mins Ant and HalfCheetah).

Check out our arXiv paper for more details!

Installation

PyPI

EnvPool is currently hosted on PyPI. It supports Python 3.11-3.14 on Linux, macOS, and Windows.

You can simply install EnvPool with the following command:

$ pip install envpool

After installation, open a Python console and type

import envpool
print(envpool.__version__)

If no error occurs, you have successfully installed EnvPool.

Platform notes:

• Linux Procgen wheels intentionally do not vendor Qt. If making a Procgen environment reports missing libQt5Core.so.5 or libQt5Gui.so.5, install the system Qt 5 runtime, for example with apt install qtbase5-dev or dnf install qt5-qtbase.
• Windows Procgen wheels bundle the required Qt runtime DLLs (Qt5Core.dll and Qt5Gui.dll) next to the extension module.
• Windows source/release CI validates MuJoCo rendering with Mesa software OpenGL. To reproduce that setup locally, point ENVPOOL_DLL_DIR at a Mesa DLL directory and set GALLIUM_DRIVER=llvmpipe plus MESA_GL_VERSION_OVERRIDE=4.5COMPAT.
• Building from source still requires platform-local build dependencies, including Qt 5. The full per-platform setup is documented in Build From Source.

From Source

Please refer to the guideline.

Documentation

The tutorials and API documentation are hosted on envpool.readthedocs.io.

The example scripts are under examples/ folder; benchmark scripts are under benchmark/ folder.

Benchmark Results

The historical benchmark tables below were produced with ALE Atari environment PongNoFrameskip-v4 (with environment wrappers from OpenAI Baselines) and MuJoCo environment Ant-v3 on different hardware setups, including a TPUv3-8 virtual machine (VM) of 96 CPU cores and 2 NUMA nodes, and an NVIDIA DGX-A100 of 256 CPU cores with 8 NUMA nodes. The current scripts under benchmark/ use Gymnasium's ALE/Pong-v5 and Ant-v5. Baselines include 1) naive Python for-loop; 2) the most popular RL environment parallelization execution by Python subprocess, e.g., gym.vector_env; 3) to our knowledge, the fastest RL environment executor Sample Factory before EnvPool.

We report EnvPool performance with sync mode, async mode, and NUMA + async mode, compared with the baselines on different number of workers (i.e., number of CPU cores). As we can see from the results, EnvPool achieves significant improvements over the baselines on all settings. On the high-end setup, EnvPool achieves 1 Million frames per second with Atari and 3 Million frames per second with MuJoCo on 256 CPU cores, which is 14.9x / 19.6x of the gym.vector_env baseline. On a typical PC setup with 12 CPU cores, EnvPool's throughput is 3.1x / 2.9x of gym.vector_env.

Atari Highest FPS	Laptop (12)	Workstation (32)	TPU-VM (96)	DGX-A100 (256)
For-loop	4,893	7,914	3,993	4,640
Subprocess	15,863	47,699	46,910	71,943
Sample-Factory	28,216	138,847	222,327	707,494
EnvPool (sync)	37,396	133,824	170,380	427,851
EnvPool (async)	49,439	200,428	359,559	891,286
EnvPool (numa+async)	/	/	373,169	1,069,922

MuJoCo Highest FPS	Laptop (12)	Workstation (32)	TPU-VM (96)	DGX-A100 (256)
For-loop	12,861	20,298	10,474	11,569
Subprocess	36,586	105,432	87,403	163,656
Sample-Factory	62,510	309,264	461,515	1,573,262
EnvPool (sync)	66,622	380,950	296,681	949,787
EnvPool (async)	105,126	582,446	887,540	2,363,864
EnvPool (numa+async)	/	/	896,830	3,134,287

Please refer to the benchmark page for more details.

API Usage

The following content shows both synchronous and asynchronous API usage of EnvPool. You can also run the full script at examples/env_step.py

Synchronous API

import envpool
import numpy as np

# make Gymnasium env
env = envpool.make("Pong-v5", env_type="gymnasium", num_envs=100)
# or use envpool.make_gymnasium(...)
obs = env.reset()  # should be (100, 4, 84, 84)
act = np.zeros(100, dtype=int)
obs, rew, term, trunc, info = env.step(act)

Under the synchronous mode, envpool closely resembles Gymnasium and dm_env. It has the reset and step functions with the same meaning. However, there is one exception in envpool: batch interaction is the default. Therefore, during the creation of the envpool, there is a num_envs argument that denotes how many envs you like to run in parallel.

env = envpool.make("Pong-v5", env_type="gymnasium", num_envs=100)

The first dimension of action passed to the step function should equal num_envs.

act = np.zeros(100, dtype=int)

You don't need to manually reset one environment when any of done is true; instead, all envs in envpool have enabled auto-reset by default.

Asynchronous API

import envpool
import numpy as np

# make asynchronous
num_envs = 64
batch_size = 16
env = envpool.make(
    "Pong-v5", env_type="gymnasium", num_envs=num_envs, batch_size=batch_size
)
action_num = env.action_space.n
env.async_reset()  # send the initial reset signal to all envs
while True:
    obs, rew, term, trunc, info = env.recv()
    env_id = info["env_id"]
    action = np.random.randint(action_num, size=batch_size)
    env.send(action, env_id)

In the asynchronous mode, the step function is split into two parts: the send/recv functions. send takes two arguments, a batch of action, and the corresponding env_id that each action should be sent to. Unlike step, send does not wait for the envs to execute and return the next state, it returns immediately after the actions are fed to the envs. (The reason why it is called async mode).

env.send(action, env_id)

To get the "next states", we need to call the recv function. However, recv does not guarantee that you will get back the "next states" of the envs you just called send on. Instead, whatever envs finishes execution gets recved first.

state = env.recv()

Besides num_envs, there is one more argument batch_size. While num_envs defines how many envs in total are managed by the envpool, batch_size specifies the number of envs involved each time we interact with envpool. e.g. There are 64 envs executing in the envpool, send and recv each time interacts with a batch of 16 envs.

envpool.make("Pong-v5", env_type="gymnasium", num_envs=64, batch_size=16)

There are other configurable arguments with envpool.make; please check out EnvPool Python interface introduction.

Rendering

EnvPool exposes rendering through the Python wrapper. Create the env with render_mode="rgb_array" to get batched RGB output, or render_mode="human" to display a single env through OpenCV.

import envpool

env = envpool.make(
    "Ant-v5",
    env_type="gymnasium",
    num_envs=4,
    render_mode="rgb_array",
    render_width=480,
    render_height=480,
)
env.reset()
frames = env.render(env_ids=[0, 2])
assert frames.shape == (2, 480, 480, 3)

render() is batch-first, so even a single render keeps the batch dimension: env.render().shape == (1, H, W, 3). If env_ids is omitted, EnvPool renders render_env_id (default 0). camera_id can be overridden per call, while the output size is fixed at env creation time via render_width and render_height.

The repo test suite also exercises rendering. make bazel-test runs repeated render checks for every render-capable env family, and make release-test includes a wheel smoke that calls render() after reset(). On Windows, the MuJoCo render tests use the same ENVPOOL_DLL_DIR Mesa preload hook described above when you want software OpenGL instead of the system driver.

viewer = envpool.make(
    "WalkerWalk-v1",
    env_type="gymnasium",
    num_envs=1,
    render_mode="human",
    render_env_id=0,
)
viewer.reset()
viewer.render()

render_mode="human" returns None and currently supports a single env id per call. It also requires opencv-python to be installed.

Pixel Observations

For MuJoCo tasks, pixel observations can also be exposed directly through the regular observation API by passing from_pixels=True. This path is produced natively in C++, without routing through Python-side render().

pixels = envpool.make(
    "WalkerWalk-v1",
    env_type="gymnasium",
    num_envs=2,
    from_pixels=True,
    frame_stack=3,
    render_width=84,
    render_height=84,
)
obs, info = pixels.reset()
assert obs.shape == (2, 9, 84, 84)

Pixel observations use channel-first layout. With frame_stack=1, each environment returns (3, H, W); with frame_stack=3, EnvPool stacks frames on the channel dimension and returns (9, H, W). This matches the usual PyTorch BCHW convention directly. If render_width / render_height are omitted, EnvPool defaults them to 84.

Contributing

EnvPool is still under development. More environments will be added, and we always welcome contributions to help EnvPool better. If you would like to contribute, please check out our contribution guideline.

License

EnvPool is under Apache2 license.

Other third-party source-code and data are under their corresponding licenses.

We do not include their source code and data in this repo.

Citing EnvPool

If you find EnvPool useful, please cite it in your publications.

@inproceedings{weng2022envpool,
 author = {Weng, Jiayi and Lin, Min and Huang, Shengyi and Liu, Bo and Makoviichuk, Denys and Makoviychuk, Viktor and Liu, Zichen and Song, Yufan and Luo, Ting and Jiang, Yukun and Xu, Zhongwen and Yan, Shuicheng},
 booktitle = {Advances in Neural Information Processing Systems},
 editor = {S. Koyejo and S. Mohamed and A. Agarwal and D. Belgrave and K. Cho and A. Oh},
 pages = {22409--22421},
 publisher = {Curran Associates, Inc.},
 title = {Env{P}ool: A Highly Parallel Reinforcement Learning Environment Execution Engine},
 url = {https://proceedings.neurips.cc/paper_files/paper/2022/file/8caaf08e49ddbad6694fae067442ee21-Paper-Datasets_and_Benchmarks.pdf},
 volume = {35},
 year = {2022}
}

Disclaimer

This is not an official Sea Limited or Garena Online Private Limited product.