Open world survival environment for reinforcement learning.
If you find this code useful, please reference in your paper:
@misc{hafner2021crafter,
title = {Crafter: An Open World Survival Benchmark},
author = {Danijar Hafner},
year = {2021},
howpublished = {\url{https://github.com/danijar/crafter}},
}
Crafter is a simulated environment that tests a variety of general abilities of learning agents. It features a randomized open-ended world with image inputs where the player discovers resources and tools, all while ensuring its own survival.
- Generalization: New procedurally generated map for each episode.
- Exploration: Materials unlock new tools which in turn unlock new materials.
- Long dependencies: Gathering resources, building shelter, growing fruits.
- Partial observability: Each input image reveals only a small part of the world.
- Survival: Must find food and water, shelter to rest, defend against monsters.
- Easy to use: Pure Python, windowless rendering, few dependencies, flat action space.
Crafter has been released in its stable version v1.x.x and can now be used for research projects. The environment will not change anymore to ensure comparability. Future changes may add additional debugging features that do not affect the environment itself.
You can play the game yourself with an interactive window and keyboard input. The mapping from keys to actions, health level, and inventory state are printed to the terminal.
# Install with GUI
pip3 install pygame
pip3 install crafter
# Start the game
crafter
# Alternative way to start the game
python3 -m crafter.run_gui
The following optional command line flags are available:
Flag | Default | Description |
---|---|---|
--record <directory> |
None | Directory for recording trajectories in NPZ and MP4 formats. |
--window <width> <height> |
600 600 | Window size in pixels. |
--area <width> <height> |
64 64 | The number of grid cells of the generated world. |
--view <width> <height> |
9 9 | The number of grid cells that are visible in the images. |
--size <width> <height> |
0 0 | Render resolution; defaults to the window size. Setting this to 64 64 shows the low resolution graphics that artificial agents see. |
--wait <boolean> |
False | Pauses the game while the player chooses their action. |
--seed <integer> |
None | Determines world generation and creatures. |
Installation: pip3 install -U crafter
The environment follows the OpenAI Gym interface:
import crafter
env = crafter.Env(seed=0)
obs = env.reset()
assert obs.shape == (64, 64, 3)
done = False
while not done:
action = env.action_space.sample()
obs, reward, done, info = env.step(action)
To ensure comparability across research papers, we recommend using the environment in its default configuration. Nonetheless, the environment can be configured via its constructor:
crafter.Env(area=(64, 64), view=(9, 9), size=(64, 64), length=10000, seed=None)
Parameter | Default | Description |
---|---|---|
area |
(64, 64) |
Size of the world in grid cells. |
view |
(9, 9) |
Layout size in cells; determines view distance. |
size |
(64, 64) |
Render size of the images in pixels. |
length |
10000 |
Time limit for the episode, can be None . |
seed |
None | Interger that determines world generation and creatures. |
The reward can either be given to the agent or used as a proxy metric for evaluating unsupervised agents.
The reward is +1 when the agent unlocks a new achievement, -0.1 when its health level decreases, +0.1 when it increases, and 0 for all other time steps. The achievements are as follows:
collect_coal
collect_diamond
collect_drink
collect_iron
collect_sapling
collect_stone
collect_wood
defeat_skeleton
defeat_zombie
eat_cow
eat_plant
make_iron_pickaxe
make_iron_sword
make_stone_pickaxe
make_stone_sword
make_wood_pickaxe
make_wood_sword
place_furnace
place_plant
place_stone
place_table
wake_up
The sum of rewards per episode can range from -0.9 (losing all health without any achievements) to 22 (unlocking all achievements and keeping or restoring all health until the time limit is reached). A score of 21.1 or higher means that all achievements have been unlocked.
The episode terminates when the health points of the agent reach zero. Episodes also end when reaching a time limit, which is 10000 steps by default.
Each observation is an RGB image that shows a local view of the world around the player, as well as the life statistics and inventory state of the agent.
The action space is categorical. Each action is an integer index representing one of the possible actions:
Integer | Name | Requirement |
---|---|---|
0 | noop |
Always applicable. |
1 | move_left |
Flat ground left to the agent. |
2 | move_right |
Flat ground right to the agent. |
3 | move_up |
Flat ground above the agent. |
4 | move_down |
Flat ground below the agent. |
5 | do |
Facing creature or material and have necessary tool. |
6 | sleep |
Energy level is below maximum. |
7 | place_stone |
Stone in inventory. |
8 | place_table |
Wood in inventory. |
9 | place_furnace |
Stone in inventory. |
10 | place_plant |
Sapling in inventory. |
11 | make_wood_pickaxe |
Nearby table. Wood in inventory. |
12 | make_stone_pickaxe |
Nearby table. Wood, stone in inventory. |
13 | make_iron_pickaxe |
Nearby table, furnace. Wood, coal, iron an inventory. |
14 | make_wood_sword |
Nearby table. Wood in inventory. |
15 | make_stone_sword |
Nearby table. Wood, stone in inventory. |
16 | make_iron_sword |
Nearby table, furnace. Wood, coal, iron an inventory. |
The step function returns an info
directionary with additional information
about the environment state. It can be used for evaluation and debugging but
should not be provided to the agent. The following entries are available:
Key | Type | Description |
---|---|---|
inventory |
dict | Mapping from item names to inventory counts. |
achievements |
dict | Mapping from achievement names to their counts. |
discount |
float | 1 during the episode and 0 at the last step. |
semantic |
np.array | Categorical representation of the world. |
player_pos |
tuple | X and Y position of the player in the world. |
Crafter is designed to be challenging for current learning algorithms but not completely out of reach. To verify how challenging the environment is, we trained the DreamerV2 agent on Crafter with rewards. We recommend training for 5M environment steps and reporting the mean score. During this time, the agent makes consistent learning progress.
When training over 10 times longer, the agent also rarely unlocks all achievements during an episode, including finding a diamond. The open research challenge ahead of us is to drastically accelerate the exploration and learning progress and increase the average score by consistently unlocking all achievements.
Please open an issue on Github.