Obstacle Tower in headless-mode

[jkuypers93]

Hi, thank you for sharing your work!

I am trying to run your code for Obstacle Tower on a cluster but I keep getting timeout errors from ml-agents

File "/home/ubuntu/anaconda3/lib/python3.6/site-packages/mlagents/envs/rpc_communicator.py", line 86, in initialize
    "The Unity environment took too long to respond. Make sure that :\n"
mlagents.envs.exception.UnityTimeOutException: The Unity environment took too long to respond. Make sure that :
	 The environment does not need user interaction to launch
	 The Academy's Broadcast Hub is configured correctly
	 The Agents are linked to the appropriate Brains

I think it is due to the fact that I am running it in headless mode. Did you run into a similar issue? Any idea on how to run it in headless mode using xvfb?

Thanks in advance!

[Miffyli]

Hey! Happy to hear people are finding the code useful :)

I see OT environment has had some updates since this project. I tried installing the newest version (this functionality should be same with code here) and running it with xvfb-run -a ..., and things seem to work (first line is to disable display):

anssk@anssk-workhorse:~/temp$ export DISPLAY=""
anssk@anssk-workhorse:~/temp$ xvfb-run -a ipython3
Python 3.8.5 (default, Jul 28 2020, 12:59:40)
Type 'copyright', 'credits' or 'license' for more information
IPython 7.13.0 -- An enhanced Interactive Python. Type '?' for help.

In [1]: from obstacle_tower_env import ObstacleTowerEnv

In [2]: env = ObstacleTowerEnv()
Found path: /tmp/ml-agents-binaries/binaries/ObstacleTower-3ab155c962d77775bc1e66b866180213/ObstacleTower/obstacletower.x86_64
INFO:mlagents_envs.environment:Connected to Unity environment with package version 1.2.0-preview and communication version 1.0.0
INFO:mlagents_envs.environment:Connected new brain:
ObstacleTowerAgent-v4.1?team=0
INFO:gym_unity:1 agents within environment.

In [3]:

Without xvfb-run I am getting the same error as you are upon creating the environment.

Can you try if this works? You can add xvfb-run -a .... before running the .sh script for obstacle tower experiments.

System:

• Ubuntu 20.04 LTS
• Python 3.8 (system's default)

[jkuypers93]

Thanks for getting back to me.

That worked for me. However, my worker nodes did not seem to have access to a virtual screen, so I had to create a display in my Worker class

from pyvirtualdisplay import Display

@ray.remote
class Worker:
    def __init__(self, config, env_creator, worker_index):
        self.display = Display(visible=0, size=(800, 600), backend="xvfb")
        self.display.start()
        ...

Thanks again for your help!