Holodeck is a high-fidelity simulator for reinforcement learning built on top of Unreal Engine 4. Read the docs.

Citation:

@misc{HolodeckPCCL,
  Author = {Joshua Greaves and Max Robinson and Nick Walton and Mitchell Mortensen and Robert Pottorff and Connor Christopherson and Derek Hancock and David Wingate},
  Title = {Holodeck: A High Fidelity Simulator},
  Year = {2018},
}

Windows and Linux:

• Python 3.5 or higher
• Pip3
• At least 3gb storage

Linux:

• OpenGL version 3 or higher

To install the python bindings, simply run pip3 install holodeck

Note: for some versions of pip you may see a warning, but holodeck will be installed anyway.

For Ubuntu 16.04 and Cuda 9.0*:

1. Install nvidia-docker.
2. Pull the docker repository and run it. (replace "without" with "with" to install with the DefaultWorlds package)
3. Holodeck cannot be run with root privileges so after entering the container and before you use holodeck you must login to a normal user. The holodeckuser with password "holodeck" has been provided. This account must be used if the with-worlds version is selected.

docker pull pccl/holodeck:ubuntu16.04-without-worlds
docker run -it pccl/holodeck:ubuntu16.04-without-worlds

*For versions other than Ubuntu 16.04 and Cuda 9.0 see here..

Holodeck currently contains one package, the DefaultWorlds package. Each package in turn contains a number of worlds. Holodeck has an internal package manager for handling packages. The most important commands for managing these are as follows:

import holodeck
print(holodeck.all_packages())                 # View all packages that are available to be downloaded.
print(holodeck.installed_packages())           # View which packages are currently installed.
holodeck.install('DefaultWorlds')              # Installs the DefaultWorlds package.
print(holodeck.package_info('DefaultWorlds'))  # View information on what worlds this package
                                               # contains, and what agents are in those worlds.
holodeck.remove('DefaultWorlds')               # Removes a package.

You only need to install packages once. You should make sure to remove them with holodeck.remove(package_name) or holodeck.remove_all_packages() before removing holodeck with pip.

Holodeck's interface is designed in the same vein as OpenAI's Gym. The quickest way to get acquainted with Holodeck use is to view the example.py file. Here is a basic walkthrough of an example that runs a Holodeck world:

import holodeck
import numpy as np
env = holodeck.make("UrbanCity")    # Load the environment. This environment contains a UAV in a city.
command = np.array([0, 0, 0, 100])  # The UAV takes 3 torques and a thrust as a command.
for i in range(30):
    state, reward, terminal, info = env.step(command)  # Pass the command to the environment with step.
                                                       # This returns the state, reward, terminal and info tuple.

The state is a dictionary of sensor enum to sensor value. Reward is the reward received from the previous action, and terminal indicates whether the current state is a terminal state. Info contains additional environment specific information.

If you want to access the data of a specific sensor, it is as simple as import Sensors and retrieving the correct value from the state dictionary:

from holodeck.sensors import Sensors
print(state[Sensors.LOCATION_SENSOR])

Holodeck supports different control schemes for different agents. Currently the only agent with multiple control schemes is the UAV agent. The control scheme can be switched as follows:

from holodeck.agents import ControlSchemes

env.set_control_scheme('uav0', ControlSchemes.UAV_ROLL_PITCH_YAW_RATE_ALT)

For more control schemes, check out the docs

Holodeck supports multi-agent environments. The interface is a little different, but still very easy to use. Instead of calling step which passes a command to the main agent and ticks the game, you should call act. act supplies a command to a specific agent, but doesn't tick the game. Once all agents have received their actions, you can call tick to tick the game. After act, every time you call tick the same command will be supplied to the agent. To change the command, just call act again.

env = holodeck.make('CyberPunkCity')
env.reset()

env.act('uav0', np.array([0, 0, 0, 100]))
env.act('nav0', np.array([0, 0, 0]))
for i in range(300):
    s = env.tick()

The state returned from tick is also somewhat different. The state is now a dictionary from agent name to sensor dictionary. You can access the reward, terminal and location for the UAV as follows:

s['uav0'][Sensors.REWARD]
s['uav0'][Sensors.TERMINAL]
s['uav0'][Sensors.LOCATION_SENSOR]

• C - toggles between a directly attached camera, which allows you to see more or less what the agent sees, and relative attach, which is the default camera attachment.
• V - toggles spectator mode, which allows you detach from the agent and explore the world without affecting the agent's vision.

You can view stats by entering console commands. When an environment is running, type ~ to open the console and enter a command. A common one to use is stat FPS to display the frames per second. More commands can be found in UDK documentation.

• Agents
• Sensors
• Environment configuration
• Docs

To create custom worlds with variable start positions, number and type of agents, and different environments see the Holodeck Engine and follow the Packaging and Using Custom Worlds wiki to use Holodeck for editing worlds with the Unreal editor.

To use OpenGL3 in linux, change the argument in Holodeck.make:

from Holodeck import Holodeck
env = Holodeck.make("MazeWorld", Holodeck.GL_VERSION.OPENGL3)

Holodeck can run on headless machines with GPU accelerated rendering. This requires no extra configuration. Holodeck will automatically detect that the machine is headless and configure it's rendering process accordingly.