Obstacle Tower Environment
The Obstacle Tower is a procedurally generated environment consisting of multiple floors to be solved by a learning agent. It is designed to test learning agents abilities in computer vision, locomotion skills, high-level planning, and generalization. It combines platforming-style gameplay with puzzles and planning problems, and critically, increases in difficulty as the agent progresses.
Within each floor, the goal of the agent is to arrive at the set of stairs leading to the next level of the tower. These floors are composed of multiple rooms, each which can contain their own unique challenges. Furthermore, each floor contains a number of procedurally generated elements, such as visual appearance, puzzle configuration, and floor layout. This ensures that in order for an agent to be successful at the Obstacle Tower task, they must be able to generalize to new and unseen combinations of conditions.
To learn more, please read our AAAI Workshop paper:
The Obstacle Tower environment runs on Mac OS X, Windows, or Linux.
Python dependencies (also in setup.py):
By default, the binary will be automatically downloaded when the Obstacle Tower gym is first instantiated. The binaries for each platform can be separately downloaded at the following links.
| Platform | Download Link | | --- | --- | | Linux (x8664) | https://storage.googleapis.com/obstacle-tower-build/v4.1/obstacletowerv4.1linux.zip | | Mac OS X | https://storage.googleapis.com/obstacle-tower-build/v4.1/obstacletowerv4.1osx.zip | | Windows | https://storage.googleapis.com/obstacle-tower-build/v4.1/obstacletowerv4.1_windows.zip |
For checksums on these files, see here.
$ git clone [email protected]:Unity-Technologies/obstacle-tower-env.git $ cd obstacle-tower-env $ pip install -e .
To see an example of how to interact with the environment using the gym interface, see our Basic Usage Jupyter Notebook.
Obstacle Tower can be configured in a number of different ways to adjust the difficulty and content of the environment. This is done through the use of reset parameters, which can be set when calling
env.reset(). See here for a list of the available parameters to adjust.
It is also possible to launch the environment in "Player Mode," and directly control the agent using a keyboard. This can be done by double-clicking on the binary file. The keyboard controls are as follows:
| Keyboard Key | Action | | --- | --- | | W | Move character forward. | | S | Move character backwards. | | A | Move character left. | | D | Move character right. | | K | Rotate camera left. | | L | Rotate camera right. | | Space | Character jump. |
We provide an environment wrapper for evaluating performance of a player or agent across multiple pre-defined seeds. We provide an example implementation of evaluation on a random policy.
If you are interested in training an agent using Google's Dopamine framework and/or Google Cloud Platform, see our guide here.