Fast and scalable reinforcement learning environment for the IGLU competition at NeurIPS 2022. The env represents an embodied agent with an ability to navigate, place, and break blocks of six different colors.

IGLU is a research project aimed at bridging the gap between reinforcement learning and natural language understanding in Minecraft as a collaborative environment. It provides the RL environment where the goal of an agent is to build structures within a dedicated zone. The structures are described by natural language in the game’s chat.

Install env:

pip install git+https://github.com/iglu-contest/gridworld.git@master

Clone the repo and build local conda env:

git clone https://github.com/iglu-contest/gridworld.git
cd gridworld && conda env create -f env.yml

docker build -t gridworld -f ./docker/Dockerfile .

Note that by default IGLU env runs in headless mode. To run headed do

export IGLU_HEADLESS=0

Now, run the environment loop:

import gym
import gridworld
from gridworld.tasks import DUMMY_TASK

# create vector based env. Rendering is enabled by default. 
# To turn it off, use render=False.
env = gym.make('IGLUGridworld-v0')

# It is mandatory to task the environemnt. 
# For dummy looping, you can use a DUMMY_TASK object:
env.set_task(DUMMY_TASK)
done = False
obs = env.reset()
while not done:
  action = env.action_space.sample()
  obs, reward, done, info = env.step(action)

Two action spaces are available in Gridworld:

Walking actions allow the agent to move with gravity enabled, jump, place, break, and select blocks from inventory.

env = gym.make('IGLUGridworld-v0', action_space='walking')
print(env.action_space) # Discrete(18)

The action space format is the following:

• 0 - no-op
• 1 - step forward
• 2 - step backward
• 3 - step left
• 4 - step right
• 5 - jump
• 6-11 - inventory select
• 12 - move the camera left
• 13 - move the camera right
• 14 - move the camera up
• 15 - move the camera down
• 16 - break block
• 17 - place block

For each movement action, the agent steps for about 0.25 of one block. Camera movement changes each angle by 5 degrees.

Flying actions allow the agent to fly freely within the building zone. Placement actions are the same and movement is specified by a continuous vector.

env = gym.make('IGLUGridworld-v0', action_space='flying')

Action space format:

Dict(
  movement: Box(low=-1, high=1, shape=(3,)),
  camera: Box(low=-5, high=5, shape=(2,)),
  inventory: Discrete(7),
  placement: Discrete(3)
)

Observation space format:

Dict(
  inventory: Box(low=0, high=20, shape=(6,)),
  compass: Box(low=-180, high=180, shape=(1,)),
  dialog: String(),
  pov: Box(low=0, high=255, shape=(64, 64, 3))
)

Here, inventory indicates the total blocks available to the agent (per color). The compass component shows the angle between the agent's yaw angle and the North direction. The dialog value is the full previous dialog and the most recent instruction to execute. Finally, pov is an ego-centric image of the agent's observation of the world. Note that this space will be used during the evaluation. However, it is possible to access other fields of the environment, for example, during training. The vector_state=True passed as keyword argument in gym.make will return, in addition to previous fields,

agentPos: Box(low=[-8, -2, -8, -90, 0], high=[8, 12, 8, 90, 360], shape=(5,)),
grid: Box(low=-1, high=7, shape=(9, 11, 11))

It is also possible to make a target grid a part of the observation space. To do that, pass target_in_obs=True to gym.make. This will add another key to the observation space with the same structure as the grid component. The name of a new component is target_grid. This part of the space remains fixed within an episode.

Each step, the reward is calculated based on the similarity between the so far built grid and the target grid. The reward is determined regardless of global spatial position of currently placed blocks, it only takes into account how much the built blocks are similar to the target structure. To make it possible, at each step we calculate the intersection between the built and the target structures for each spatial translation within the horizontal plane and rotation around the vertical axis. Then we take the maximal intersection value among all translation and rotations. To calculate the reward, we compare the maximal intersection size from the current step with the one from the previous step. We reward the agent with 2 for the increase of the maximal intersection size, with -2 for the decrease of the maximal intersection size, and with 1/-1 for removing/placing a block without a change of the maximal intersection size. A visual example is shown below.

Specifically, we run the code that is equivalent to the following one:

def maximal_intersection(grid, target_grid):
  """
  Args:
    grid (np.ndarray[Y, X, Z]): numpy array snapshot of a built structure
    target_grid (np.ndarray[Y, X, Z]): numpy array snapshot of the target structure
  """
  maximum = 0
  # iterate over orthogonal rotations
  for i in range(4):
    # iterate over translations
    for dx in range(-X, X + 1):
      for dz in range(-Z, Z + 1):
        shifted_grid = translate(grid, dx, dz)
        intersection = np.sum( (shifted_grid == target) & (target != 0) )
        maximum = max(maximum, intersection)
    grid = rotate_y_axis(grid)
  return maximum

In practice, a more optimized version is used. The reward is then calculated based on the temporal difference between maximal intersection of the two consecutive grids. Formally, suppose grids[t] is a built structure at timestep t. The reward is then calculated as:

def calc_reward(prev_grid, grid, target_grid, , right_scale=2, wrong_scale=1):
  prev_max_int = maximal_intersection(prev_grid, target_grid)
  max_int = maximal_intersection(grid, target_grid)
  diff = max_int - prev_max_int
  prev_grid_size = num_blocks(prev_grid)
  grid_size = num_blocks(grid)
  if diff == 0:
    return wrong_scale * np.sign(grid_size - prev_grid_size)
  else:
    return right_scale * np.sign(diff)

In other words, if a recently placed block strictly increases or decreases the maximal intersection, the reward is positive or negative and is equal to +/-right_scale. Otherwise, its absolute value is equal to wrong_scale and the sign is positive if a block was removed or negative if added. Values right_scale and wrong_scale can be passed to gym.make as environment kwargs. Finally, the maximal_intersection includes heavy computations that slow down the environment. They can be simplified by disabling rotational/translational invariance at the cost of much more sparse reward. To do that, pass invariant=False to a corresponding Task object (see Dataset section for reference).

By default, the environment requires a task object to run. IGLU dataset provides a convenient loader for RL tasks. Here is an example of how to use it:

import gym
from gridworld.data import IGLUDataset

dataset = IGLUDataset(dataset_version='v0.1.0-rc1') 
# leave dataset_version empty to access the most recent version of the dataset.

env = gym.make('IGLUGridworld-v0')
env.set_task_generator(dataset)

In this example, we download the dataset of tasks for RL env. Internally, on each .reset() of the env, the dataset samples a random task (inside its own .reset() method) and makes it active in the env. The Task object is responsible for calculating the reward, providing the text part of the observation, and determining if the episode has ended.

The structure of the IGLU dataset is following. The dataset consists of structures that represent overall collaboration goals. For each structure, we have several collaboration sessions that pair architects with builders to build each particular structure. Each session consists of a sequence of "turns". Each turn represents an atomic instruction and corresponding changes of the blocks in the world. The structure of a Task object is following:

• target_grid - target blocks configuration that needs to be built
• starting_grid - optional, blocks for the environment to begin the episode with.
• chat - full conversation between the architect and builder, including the most recent instruction
• last_instruction - last utterance of the architect

Sometimes, the instructions can be ambiguous and the builder asks a clarifying question which the architect answers. In the latter case, last_instruction will contain three utterances: an instruction, a clarifying question, and an answer to that question. Otherwise, last_instruction is just one utterance of the architect.

To represent collaboration sessions, the Subtasks class is used. This class represents a sequence of dialog utterances and their corresponding goals (each of which is a partially completed structure). On .reset() call, it picks a random turn and returns a Task object, where starting and target grids are consecutive partial structures and the dialog contains all utterances up until the one corresponding to the target grid.

In the example above, the dataset object is structured as follows:

# .tasks is a dict mapping from structure to a list of sessions of interaction
dataset.tasks 
# each value contains a list corresponding to collaboration sessions.
dataset.tasks['c73']
# Each element of this list is an instance of `Subtasks` class
dataset.tasks['c73'][0]

The .reset() method of IGLUDataset does effectively the following:

def reset(dataset):
  task_id = random.choice(dataset.tasks.keys())
  session = random.choice(dataset.tasks[task_id])
  subtask = session.reset() # Task object is returned
  return subtask

This behavior can be customized simply by overriding the reset method in a subclass:

import gym
from gridworld.data import IGLUDataset

class MyDataset(IGLUDataset):
  def __init__(self, *args, **kwargs):
    super().__init__(*args, **kwargs)
    self.my_task_id = 'c73'
    self.my_session = 0
  
  def reset(self):
    return self.tasks[self.my_task_id][self.my_session].reset()

env = gym.make('IGLUGridworld-v0')
my_dataset = MyDataset(dataset_version='v0.1.0-rc1')
env.set_task_generator(my_dataset)
# do training/sampling

On the first creation, the dataset is downloaded and parsed automatically. Below you will find the structure of the dataset:

dialogs.csv
builder-data/
  ...
  1-c118/ # session id - structure_id
    step-2
  ...
  9-c118/
    step-2
    step-4
    step-6
  1-c120/
    step-2
  ...
  23-c126/
    step-2
    step-4
    step-6
    step-8

Here, dialog.csv contains the utterances of architects and builders solving different tasks in different sessions. The builder-data/ directory contains builder behavior recorded by the voxel.js engine. Right now we extract only the resulting grids and use them as targets.