Great installation guide for universe on Ubuntu


After spending about a week and a half trying to install universe per instructions available on I still couldn’t get universe fully working. With a recent web search I came across the following “JUSTIN’S BLOG” “OpenAI – Universe Installation Guide Ubuntu 16.04”:

I’ve been working with an Ubuntu 16.04 VM, so I created a new virgin user on my VM and did a new universe installation using Justin’s install procedure. I followed it almost to the letter, with one exception noted later. After a couple hours, working slowly and methodically, I had what so far seems to be a fully working universe installation. I can run the universe-starting-agent and see all the Tensorflow stuff in my browser. Lastly I don’t know if Chromium is a better choice than Firefox to view Tensorflow but I’m using Chromium now per Justin’s instructions.

Justin’s guide is a relatively simple step-by-step and nearly complete procedure. The only step I had problems with was his instructions to get Tensorflow from I kept getting the error that it wouldn’t run on my platform. Possibly something to do with me having Python 3.6 and not 3.4 or 3.5. I didn’t want to mess with installing multiple Pythons and finally installed Tensorflow according to , Basically just “pip install tensorflow” with the conda universe source activated, and with user Home my current directory. Every other step was per Justin’s guide, and in the order suggested. Also I chose the CPU version of Tensorflow to keep it simple. If I later want to explore using GPU I’ll do it with a separate user or VM installation.

The non-specific instructions on the Github getting started left me too much room to hurt myself. I had wound up installing Docker CE and Miniconda3 for instance. And I don’t even recall what order I wound up installing/reinstalling the universe components. If anyone is still struggling with getting their universe install working I can highly recommend starting over with a clean environment (eg. new Ubuntu user) and proceeding with the guide on Justin’s Blog.


Super glad you found it helpful!


You may recall I recommended the Universe install procedure from your Blog a few months back. I’ve used it about 5 times setting up multiple test machines. In fact I’m using it right at this instant to install CUDA and rebuild Tensorflow on my only Nvidia machine. I hope the forum is using it. I still see the same topics on install headaches cropping up daily on the forum, which they shouldn’t if they would just follow your procedure.

On another topic I’m looking for reviewers for a Pong parallel processing training enhancement eval and GIST I just uploaded - . The Pong results are almost an afterthought, other than the resulting model is moderately effective… The real purpose was to speed training with parallel processing, and to explore game screen pre-processing (eg. add filtering or for custom rewarder). If its not up your alley, or you’re too busy no prob. If not maybe you know someone it might interest. - Thanks


Thank you very much for a very clear and crisp tutorial on installing and running Ubuntu.Installation procedure explained clearly.


Glad it was helpful but giving credit where its due Justin’s procedure was the rock solid foundation. I only added something related to the changing landscape since Justin first published his procedure.


I might suggest adding one additional step to Justin’s procedure. This would be to manually download a few docker images just after installing docker. This isn’t essential because when you run an application that needs a docker image it will download it for you (assuming you have a network connection at the time). This works fine when you run the docker Hello World app because the docker hello-world image is only about 1.8KB so little or no download delay is noticed. But the Pong env apps depend on image and the Dusk Drive env depends on docker image. Both of these at about 1.8GB and can each take several minutes to download depending on your network speed. In fact I have two versions of the …gym.core on my systems. The older version has the PongDeterministic env needed for the universe-starter-agent, while the latest one doesn’t. seem to. Below are some docker commands that can be run from an Ubuntu terminal. I’m listing some of these from memory as I don’t have a machine ready to run all them just now:

See the docker images on your machine:
my_user@my_machine$ docker images
my_user@my_machine$ docker image list

To download the game images:
my_user@my_machine$ docker pull
my_user@my_machine$ docker pull
my_user@my_machine$ docker pull

These should be downloaded sequentially as they share many common components and running them in parallel could download the common components multiple times.

As mentioned above these steps aren’t strictly essential but they may help avoid user confusion about very, very slow env startup the first time, and might give the newbie some insight into the interplay between universe and docker.


Does anyone know a guide which shows how to make a simple and “clean” setup, maybe installing everything in a docker environment rather than using anaconda? I’m preferably interested in a more “Top-Level” article which explains how to chose the best overall setup (docker vs anaconda, tensorflow, etc.) rather than step-step guide? Thanks