/crafter

Benchmarking the Spectrum of Agent Capabilities

Primary LanguagePythonMIT LicenseMIT

Status: Stable release

PyPI

Crafter

Open world survival game for evaluating a wide range of agent abilities within a single environment.

Crafter Terrain

Overview

Crafter features randomly generated 2D worlds where the player needs to forage for food and water, find shelter to sleep, defend against monsters, collect materials, and build tools. Crafter aims to be a fruitful benchmark for reinforcement learning by focusing on the following design goals:

  • Research challenges: Crafter poses substantial challenges to current methods, evaluating strong generalization, wide and deep exploration, representation learning, and long-term reasoning and credit assignment.

  • Meaningful evaluation: Agents are evaluated by semantically meaningful achievements that can be unlocked in each episode, offering insights into the ability spectrum of both reward agents and unsupervised agents.

  • Iteration speed: Crafter evaluates many agent abilities within a single env, vastly reducing the computational requirements over benchmarks suites that require training on many separate envs from scratch.

See the research paper to find out more: Benchmarking the Spectrum of Agent Capabilities

@article{hafner2021crafter,
  title={Benchmarking the Spectrum of Agent Capabilities},
  author={Danijar Hafner},
  year={2021},
  journal={arXiv preprint arXiv:2109.06780},
}

Play Yourself

python3 -m pip install crafter  # Install Crafter
python3 -m pip install pygame   # Needed for human interface
python3 -m crafter.run_gui      # Start the game
Keyboard mapping (click to expand)
Key Action
WASD Move around
SPACE Collect material, drink from lake, hit creature
TAB Sleep
T Place a table
R Place a rock
F Place a furnace
P Place a plant
1 Craft a wood pickaxe
2 Craft a stone pickaxe
3 Craft an iron pickaxe
4 Craft a wood sword
5 Craft a stone sword
6 Craft an iron sword

Crafter Video

Interface

To install Crafter, run pip3 install crafter. The environment follows the OpenAI Gym interface. Observations are images of size (64, 64, 3) and outputs are one of 17 categorical actions.

import gym
import crafter

env = gym.make('CrafterReward-v1')  # Or CrafterNoReward-v1
env = crafter.Recorder(
  env, './path/to/logdir',
  save_stats=True,
  save_video=False,
  save_episode=False,
)

obs = env.reset()
done = False
while not done:
  action = env.action_space.sample()
  obs, reward, done, info = env.step(action)

Evaluation

Agents are allowed a budget of 1M environmnent steps and are evaluated by their success rates of the 22 achievements and by their geometric mean score. Example scripts for computing these are included in the analysis directory of the repository.

  • Reward: The sparse reward is +1 for unlocking an achievement during the episode and -0.1 or +0.1 for lost or regenerated health points. Results should be reported not as reward but as success rates and score.

  • Success rates: The success rates of the 22 achievemnts are computed as the percentage across all training episodes in which the achievement was unlocked, allowing insights into the ability spectrum of an agent.

  • Crafter score: The score is the geometric mean of success rates, so that improvements on difficult achievements contribute more than improvements on achievements with already high success rates.

Scoreboards

Please create a pull request if you would like to add your or another algorithm to the scoreboards. For the reinforcement learning and unsupervised agents categories, the interaction budget is 1M. The external knowledge category is defined more broadly.

Reinforcement Learning

Algorithm Score (%) Reward Open Source
Curious Replay 19.4±1.6 - AutonomousAgentsLab/cr-dv3
PPO (ResNet) 15.6±1.6 10.3±0.5 snu-mllab/Achievement-Distillation
DreamerV3 14.5±1.6 11.7±1.9 danijar/dreamerv3
LSTM-SPCNN 12.1±0.8 astanic/crafter-ood
EDE 11.7±1.0 yidingjiang/ede
OC-SA 11.1±0.7 astanic/crafter-ood
DreamerV2 10.0±1.2 9.0±1.7 danijar/dreamerv2
PPO 4.6±0.3 4.2±1.2 DLR-RM/stable-baselines3
Rainbow 4.3±0.2 6.0±1.3 Kaixhin/Rainbow

Unsupervised Agents

Algorithm Score (%) Reward Open Source
Plan2Explore 2.1±0.1 2.1±1.5 danijar/dreamerv2
RND 2.0±0.1 0.7±1.3 alirezakazemipour/PPO-RND
Random 1.6±0.0 2.1±1.3

External Knowledge

Algorithm Score (%) Reward Uses Interaction Open Source
Human 50.5±6.8 14.3±2.3 Life experience 0 crafter_human_dataset
SPRING 27.3±1.2 12.3±0.7 LLM, scene description, Crafter paper 0
Achievement Distillation 21.8±1.4 12.6±0.3 Reward structure 1M snu-mllab/Achievement-Distillation
ELLM 6.0±0.4 LLM, scene description 5M

Baselines

Baseline scores of various agents are available for Crafter, both with and without rewards. The scores are available in JSON format in the scores directory of the repository. For comparison, the score of human expert players is 50.5%. The baseline implementations are available as a separate repository.

Questions

Please open an issue on Github.