nle 1.3.0

The NetHack Learning Environment (NLE): a reinforcement learning environment based on NetHack

---

The NetHack Learning Environment (NLE) is a Reinforcement Learning environment originally presented at NeurIPS 2020. This version of NLE is based on NetHack 3.6.7 and designed to provide a standard RL interface to the game, and comes with tasks that function as a first step to evaluate agents on this new environment.

NetHack is one of the oldest and arguably most impactful videogames in history, as well as being one of the hardest roguelikes currently being played by humans. It is procedurally generated, rich in entities and dynamics, and overall an extremely challenging environment for current state-of-the-art RL agents, while being much cheaper to run compared to other challenging testbeds. Through NLE, we wish to establish NetHack as one of the next challenges for research in decision making and machine learning.

You can read more about NLE in the NeurIPS 2020 paper, and about NetHack in its original README, at nethack.org, and on the NetHack wiki.

This version of NLE uses the Farama Organisation Gymnasium Environment APIs.

Getting started

Starting with NLE environments is extremely simple, provided one is familiar with other gym (or Gynmasium) / RL environments.

Installation

NLE requires python>=3.10, cmake>=3.28 to be installed and available both when building the package, and at runtime.

On MacOS, one can use Homebrew as follows:

$ brew install cmake

On a plain Ubuntu 18.04 distribution, cmake and other dependencies can be installed by doing:

# Python and most build deps
$ sudo apt-get install -y build-essential autoconf libtool pkg-config \
    python3-dev python3-pip python3-numpy git flex bison libbz2-dev

# recent cmake version
$ wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | sudo apt-key add -
$ sudo apt-add-repository 'deb https://apt.kitware.com/ubuntu/ bionic main'
$ sudo apt-get update && apt-get --allow-unauthenticated install -y \
    cmake \
    kitware-archive-keyring

Afterwards it's a matter of setting up your environment. We advise using a conda environment for this:

$ conda create -y -n nle python=3.10
$ conda activate nle
$ pip install nle

NOTE: If you want to extend / develop NLE, please install the package as follows:

$ git clone https://github.com/NetHack-LE/nle
$ pip install -e ".[dev]"
$ pre-commit install

Docker

We have provided some docker images. Please see the relevant README.

Trying it out

After installation, one can try out any of the provided tasks as follows:

>>> import gymnasium as gym
>>> import nle
>>> env = gym.make("NetHackScore-v0")
>>> env.reset()  # each reset generates a new dungeon
>>> env.step(1)  # move agent '@' north
>>> env.render()

NLE also comes with a few scripts that allow to get some environment rollouts, and play with the action space:

# Play NetHackStaircase-v0 as a human
$ python -m nle.scripts.play

# Use a random agent
$ python -m nle.scripts.play --mode random

# Play the full game using directly the NetHack internal interface
# (Useful for debugging outside of the gymnasium environment)
$ python -m nle.scripts.play --env NetHackScore-v0 # works with random agent too

# See all the options
$ python -m nle.scripts.play --help

Note that nle.scripts.play can also be run with nle-play, if the package has been properly installed.

Additionally, a TorchBeast agent is bundled in nle.agent together with a simple model to provide a starting point for experiments:

$ pip install "nle[agent]"
$ python -m nle.agent.agent --num_actors 80 --batch_size 32 --unroll_length 80 --learning_rate 0.0001 --entropy_cost 0.0001 --use_lstm --total_steps 1000000000

Plot the mean return over the last 100 episodes:

$ python -m nle.scripts.plot

                              averaged episode return

  140 +---------------------------------------------------------------------+
      |             +             +            ++-+ ++++++++++++++++++++++++|
      |             :             :          ++++++++||||||||||||||||||||||||
  120 |-+...........:.............:...+-+.++++|||||||||||||||||||||||||||||||
      |             :        +++++++++++++++||||||||||AAAAAAAAAAAAAAAAAAAAAA|
      |            +++++++++++++||||||||||||||AAAAAAAAAAAA|||||||||||||||||||
  100 |-+......+++++|+|||||||||||||||||||||||AA||||||||||||||||||||||||||||||
      |       +++|||||||||||||||AAAAAAAAAAAAAA|||||||||||+++++++++++++++++++|
      |    ++++|||||AAAAAAAAAAAAAA||||||||||||++++++++++++++-+:             |
   80 |-++++|||||AAAAAA|||||||||||||||||||||+++++-+...........:...........+-|
      | ++|||||AAA|||||||||||||||++++++++++++-+ :             :             |
   60 |++||AAAAA|||||+++++++++++++-+............:.............:...........+-|
      |++|AA||||++++++-|-+        :             :             :             |
      |+|AA|||+++-+ :             :             :             :             |
   40 |+|A+++++-+...:.............:.............:.............:...........+-|
      |+AA+-+       :             :             :             :             |
      |AA-+         :             :             :             :             |
   20 |AA-+.........:.............:.............:.............:...........+-|
      |++-+         :             :             :             :             |
      |+-+          :             :             :             :             |
    0 |-+...........:.............:.............:.............:...........+-|
      |+            :             :             :             :             |
      |+            +             +             +             +             |
  -20 +---------------------------------------------------------------------+
      0           2e+08         4e+08         6e+08         8e+08         1e+09
                                       steps

NLE Language Wrapper

We thank ngoodger for implementing the NLE Language Wrapper that translates the non-language observations from NetHack tasks into similar language representations. Actions can also be optionally provided in text form which are converted to the Discrete actions of the NLE.

NetHack Learning Dataset

The NetHack Learning Dataset (NLD) code now ships with NLE, allowing users to the load large-scale datasets featured in Dungeons and Data: A Large-Scale NetHack Dataset, while also generating and loading their own datasets.

import nle.dataset as nld

if not nld.db.exists():
    nld.db.create()
    # NB: Different methods are used for data based on NLE and data from NAO.
    nld.add_nledata_directory("/path/to/nld-aa", "nld-aa-v0")
    nld.add_altorg_directory("/path/to/nld-nao", "nld-nao-v0")

dataset = nld.TtyrecDataset("nld-aa-v0", batch_size=128, ...)
for i, mb in enumerate(dataset):
    foo(mb) # etc...

For information on how to download NLD-AA and NLD-NAO, see the dataset doc here.

Otherwise checkout the tutorial Colab notebook here.

Papers using the NetHack Learning Environment

• Paglieri et al. BALROG: Benchmarking Agentic LLM and VLM Reasoning On Games (UCL, IDEAS NCBR, NYU, Oxford, Anthropic, ICLR 2025)
• Klissarov et al. MaestroMotif: Skill Design from Artificial Intelligence Feedback (Mila, FAIR, UT Austin, Alberta, Amii, ICLR 2025)
• Klissarov et al. Motif: Intrinsic Motivation from Artificial Intelligence Feedback (Mila, FAIR, UT Austin, ICLR 2024)
• Izumiya and Simo-Serra Inventory Management with Attention-Based Meta Actions (Waseda University, CoG 2021).
• Samvelyan et al. MiniHack the Planet: A Sandbox for Open-Ended Reinforcement Learning Research (FAIR, UCL, Oxford, NeurIPS 2021).
• Zhang et al. BeBold: Exploration Beyond the Boundary of Explored Regions (Berkley, FAIR, Dec 2020).
• Küttler et al. The NetHack Learning Environment (FAIR, Oxford, NYU, Imperial, UCL, NeurIPS 2020).

Open a pull request to add papers.

Contributing

We welcome contributions to NLE. If you are interested in contributing please see this document.

Architecture

NLE is direct fork of NetHack and therefore contains code that operates on many different levels of abstraction. This ranges from low-level game logic, to the higher-level administration of repeated nethack games, and finally to binding of these games to Python gymnasium environment.

If you want to learn more about the architecture of nle and how it works under the hood, checkout the architecture document. This may be a useful starting point for anyone looking to contribute to the lower level elements of NLE.

Related Environments

• gym_nethack
• rogueinabox
• rogue-gym
• MiniGrid
• CoinRun
• MineRL
• Project Malmo
• OpenAI Procgen Benchmark
• Obstacle Tower

Interview about the environment with Weights&Biases

Facebook AI Research’s Tim & Heiner on democratizing reinforcement learning research.

Citation

If you use NLE in any of your work, please cite:

@inproceedings{kuettler2020nethack,
  author    = {Heinrich K{\"{u}}ttler and
               Nantas Nardelli and
               Alexander H. Miller and
               Roberta Raileanu and
               Marco Selvatici and
               Edward Grefenstette and
               Tim Rockt{\"{a}}schel},
  title     = {{The NetHack Learning Environment}},
  booktitle = {Proceedings of the Conference on Neural Information Processing Systems (NeurIPS)},
  year      = {2020},
}

If you use NLD or the datasets in any of your work, please cite:

@article{hambro2022dungeons,
  title={Dungeons and Data: A Large-Scale NetHack Dataset},
  author={Hambro, Eric and Raileanu, Roberta and Rothermel, Danielle and Mella, Vegard and Rockt{\"a}schel, Tim and K{\"u}ttler, Heinrich and Murray, Naila},
  journal={Advances in Neural Information Processing Systems},
  volume={35},
  pages={24864--24878},
  year={2022}
}