nes-py

nes-py is a native NES emulator with a modern Gymnasium interface for reinforcement learning, scripted gameplay, emulator experimentation, and custom environment design. It runs on macOS, Linux, and Windows, builds on the SimpleNES emulator, and currently supports CPython 3.13 and 3.14 in CI.

Bring your own legally obtained .nes ROM, then use nes-py directly as a Gymnasium environment, from the bundled command-line player, or through one of the game-specific environment packages in the same ecosystem.

Highlights

• Gymnasium-native reset, step, render-mode, and seeding semantics.
• Native C++ emulator core packaged as the nes_py._native extension.
• Manual, random, windowed, and headless command-line play modes.
• NES joypad wrappers for compact reinforcement-learning action spaces.
• Mapper support for common cartridges including NROM, MMC1, UxROM, CNROM, MMC3, MMC5, AxROM, MMC2, and Sunsoft FME-7.
• Cross-platform wheels and source distributions published through PyPI trusted publishing.

Ecosystem

Use nes-py directly for arbitrary NES ROMs, or start from one of the focused Gymnasium environment packages:

• gym-super-mario-bros for Super Mario Bros.
• gym-tetris for Tetris.
• gym-zelda-1 for The Legend of Zelda.

Installation

Install nes-py from PyPI:

pip install nes-py

Python 3.13 or newer is required. The supported CI wheel targets are CPython 3.13 and 3.14.

Binary wheels are published for Linux, macOS, and Windows on those supported Python versions. If pip cannot find a compatible wheel for your interpreter or platform, it will fall back to a source build and therefore needs a working native C++ toolchain in the active environment.

Debian

Make sure you have the clang++ compiler installed:

sudo apt-get install clang

Windows

You'll need to install the Visual-Studio 17.0 tools for Windows installation. The Visual Studio Community package provides these tools for free.

Native Runtime Troubleshooting

nes-py ships a native extension, so import-time loader failures usually point to a compiler-runtime mismatch rather than a Python API bug.

• On Linux, errors mentioning GLIBCXX_* not found mean the active libstdc++.so.6 is older than the one expected by the installed wheel or build artifacts. Update the environment's C++ runtime, use a newer distribution/toolchain, or rebuild from source inside the target environment with pip install --no-binary nes-py nes-py.
• On Windows, build failures usually mean the MSVC C++ build tools are missing from the selected Python environment. Install Visual Studio Build Tools 2022 or the full Visual Studio Community package with the desktop C++ workload.
• If you are using conda, venv, or another isolated environment manager, make sure the compiler runtime loaded at import time matches the Python environment where nes-py was installed.

Usage

Command Line

Launch an interactive emulator session with a local ROM:

python3 -m nes_py.play --rom <path_to_rom>

The installed console script exposes the same interface:

nes_py --rom <path_to_rom>

Print the command-line help with:

python3 -m nes_py.play -h

The play command supports keyboard controls and random controls. Random play is handy for smoke tests and can run with or without a graphical window:

python3 -m nes_py.play --rom <path_to_rom> --mode random --steps 500
python3 -m nes_py.play --rom <path_to_rom> --mode random --steps 500 --no-render

Python API

Construct the environment with the desired Gymnasium render mode, seed through reset, and handle the separated termination and truncation flags:

from nes_py.nes_env import NESEnv

env = NESEnv("<path_to_rom>", render_mode="rgb_array")
observation, info = env.reset(seed=123)
terminated = False
truncated = False

while not (terminated or truncated):
    action = env.action_space.sample()
    observation, reward, terminated, truncated, info = env.step(action)
    frame = env.render()

env.close()

ML Observation Helpers

NESEnv.step and render in rgb_array mode keep returning the default (240, 256, 3) uint8 RGB screen view. That view is zero-copy, but it is strided over the native 32-bit screen buffer. Training loops that need a C-contiguous RGB frame or a grayscale frame can opt into explicit copy helpers and reuse output buffers:

import numpy as np

from nes_py.nes_env import NESEnv
from nes_py.nes_env import SCREEN_SHAPE_24_BIT
from nes_py.nes_env import SCREEN_SHAPE_GRAYSCALE

env = NESEnv("<path_to_rom>")
rgb = np.empty(SCREEN_SHAPE_24_BIT, dtype=np.uint8)
gray = np.empty(SCREEN_SHAPE_GRAYSCALE, dtype=np.uint8)

observation, info = env.reset()
contiguous_rgb = env.observation("rgb_array_contiguous", output=rgb)
grayscale = env.observation("grayscale", output=gray)

The helpers are intended for measured ML pipelines, not as a replacement for Gymnasium's default observation contract. Benchmark local workloads with:

python3 -m nes_py.speedtest --rom <path_to_rom> --observation-profile --no-progress

Vector, RAM, and Snapshot Helpers

Same-ROM training loops can opt into a native vector emulator that batches controller writes, frame stepping, observation copies, RAM readback, and per-slot resets without moving game-specific reward or info logic into nes-py:

import numpy as np

from nes_py.vector_env import VectorNESEmulator

vector = VectorNESEmulator("<path_to_rom>", 4)
vector.reset()
screens = vector.step(np.array([0, 1, 2, 3], dtype=np.uint8))
gray = vector.observation("grayscale")
ram_values = vector.ram_values((0x0000, (0x0001, 2, "little")))
snapshot = vector.dump_state(0)
vector.load_state(0, snapshot)
vector.close()

Scalar NESEnv also exposes ram_values(specs, output=None), dump_state(), and load_state(snapshot). Snapshots are opaque same-process checkpoint objects, not a stable cross-version save-state format.

Design notes and benchmark decisions live in:

• docs/vector-native-emulator.md
• docs/native-batch-ram-info-reads.md
• docs/vector-throughput-instrumentation.md
• docs/explicit-state-snapshot-api.md

Controls

Keyboard Key	NES Joypad
W	Up
A	Left
S	Down
D	Right
O	A
P	B
Enter	Start
Space	Select

Parallelism Caveats

Both the threading and multiprocessing packages are supported by nes-py. The rendering caveats only apply to windowed human rendering:

1. rgb_array rendering is supported from threading.Thread and multiprocessing.Process instances.
2. human rendering is not supported from instances of threading.Thread; it must run on the process's main Python thread.
3. human rendering is supported from instances of multiprocessing.Process, but the viewer must be created in the process that owns the render call. Importing nes-py or nes_py.play in a parent process does not initialize the windowing backend.

Development

To design a custom environment, introduce new emulator features, or fix a bug, start with the Wiki. It includes:

• setting up the development environment
• designing environments based on the NESEnv class
• reference material for the NESEnv API
• documentation for the nes_py.wrappers module

Project metadata, runtime dependencies, release extras, console scripts, and package discovery are configured in pyproject.toml. The native emulator source tree lives under nes_emu, with public and internal headers below nes_emu/include/nes_emu and C++ sources below nes_emu/src/nes_emu. CMake builds those sources into the nes_py._native extension through scikit-build-core. The runtime binding imports nes_py._native directly; the old ctypes shared-library discovery path is no longer used. For local development, install the package in editable mode, run the Python test suite, and build distributions through the standard PEP 517 frontend:

python -m pip install --upgrade pip build
python -m pip install --editable . --config-settings=editable.mode=inplace
python -m unittest discover .
./main.sh clean
python -m build

Native emulator internals use opt-in CMake test and benchmark targets so normal Python installs do not fetch C++ test dependencies:

cmake -S . -B build/nes-emu-debug -DCMAKE_BUILD_TYPE=Debug -DNES_EMU_BUILD_TESTS=ON
cmake --build build/nes-emu-debug --target nes_emu_tests
ctest --test-dir build/nes-emu-debug --output-on-failure
cmake -S . -B build/nes-emu-release -DCMAKE_BUILD_TYPE=Release -DNES_EMU_BUILD_BENCHMARKS=ON
cmake --build build/nes-emu-release --target nes_emu_benchmarks

PyPI releases are published by the Publish to PyPI GitHub Actions workflow through PyPI trusted publishing, not by local twine credentials. Configure the PyPI project publisher with owner Kautenja, repository nes-py, workflow filename publish.yml, and environment pypi. Then create a GitHub release from a tag matching pyproject.toml's version, with or without a leading v. The workflow builds the source distribution and CPython 3.13 and 3.14 wheels for Linux, Windows, and macOS before publishing.

Benchmarking

Developer throughput checks are available through the packaged speedtest module:

python -m nes_py.speedtest --rom nes_py/tests/games/super-mario-bros-1.nes --steps 5000

Use --json for machine-readable output. Benchmark numbers are informational and vary by machine, compiler, runner load, and display settings; they are not correctness criteria. Backup and restore stress options use explicit interval semantics, so --backup-interval 12 runs a backup at steps 12, 24, 36, and so on.

Additional profiles are available for ML and vector workflows:

python -m nes_py.speedtest --rom nes_py/tests/games/super-mario-bros-1.nes --observation-profile --json --no-progress
python -m nes_py.speedtest --rom nes_py/tests/games/super-mario-bros-1.nes --ram-profile --json --no-progress
python -m nes_py.speedtest --rom nes_py/tests/games/super-mario-bros-1.nes --vector-profile --runs 5 --env-counts 1,2,4,8,16 --instrumentation --json --no-progress

Cartridge Mapper Compatibility

nes-py supports the following cartridge mappers:

0. NROM
1. MMC1 / SxROM
2. UxROM
3. CNROM
4. MMC3 / TxROM
5. MMC5 / ExROM
6. AxROM / AOROM
7. MMC2 / PxROM
8. Sunsoft FME-7 / Sunsoft 5B

MMC3 support includes PPU A12 observations during the sprite-fetch phase, which keeps scanline IRQ split-screen effects such as Super Mario Bros. 3 status bars aligned with the rendered frame. MMC5 vertical split rendering and MMC5 pulse/PCM audio are not implemented yet. Sunsoft 5B audio register state is preserved, but expansion-audio mixing is not implemented yet. Planned mapper expansion is tracked in the umbrella repository's mapper specs.

Citation

Please cite nes-py if you use it in your research.

@misc{nes-py,
  author = {Christian Kauten},
  howpublished = {GitHub},
  title = {{NES-py}: An {NES} Emulator and {Gymnasium} Interface},
  URL = {https://github.com/Kautenja/nes-py},
  year = {2018},
}

Disclaimer

This project is provided for educational purposes only. It is not affiliated with and has not been approved by Nintendo.