Sidechannel in gym environment

Hi,

I've been moving my project from 0.10 to 0.15.0 ML-agents, and one of the key differences is that time-scaling and width/height setup has been removed from the editor, as the academy object has been removed. I now understand that you can still change these, but as a parameter when using mlagents-learn, or implement sidechannel when making UnityEnvironment.

However, when making UnityEnv, because i'm using the gym_wrapper and stable baselines, i am unable to do so. Am i missing something, can someone direct me to a place where i can read more on how to pass these arguments in a gym environment?

 

To clarify i have it running as such:
In envs\__init.py
Add:
from mlagents_envs.side_channel.engine_configuration_channel import EngineConfigurationChannel
Change:
def __init__(
self,
environment_filename: str,
dimensions: int = [], #added
timescale: int = 1, #added
worker_id: int = 0,
use_visual: bool = False,
uint8_visual: bool = False,
multiagent: bool = False,
flatten_branched: bool = False,
no_graphics: bool = False,
allow_multiple_visual_obs: bool = False,
set_config: bool = True, #added
):

base_port = 5005
if environment_filename is None:
base_port = UnityEnvironment.DEFAULT_EDITOR_PORT

#Added
channel = EngineConfigurationChannel()
if set_config==True:
channel.set_configuration_parameters(
time_scale = timescale,
width = dimensions[0],
height = dimensions[1])
#Added



self._env = UnityEnvironment(
environment_filename,
worker_id,
base_port=base_port,
no_graphics=no_graphics,
side_channels=[channel], #added
)

Was just wondering if there was an established solution other than this.

 

I ended up doing the same workaround

 

I'm trying the same work around in 0.14.1.

I changed the file ml-agents\gym-unity\gym_unity\envs\__init__.py of the gym wrapper according to your suggestion. However the following notebook still gives me errors:

1: from gym_unity.envs import UnityEnv
2: from mlagents_envs.side_channel.engine_configuration_channel import EngineConfigurationChannel
3:
4: channel1 = EngineConfigurationChannel()
5: env = UnityEnv(env_name, worker_id=1, use_visual=False, multiagent=True, side_channels=[channel1])

TypeError: __init__() got an unexpected keyword argument 'side_channels'

Could it be that I applied the changes incorrectly?
Or did I use the UnityEnv afterwards wrongly?

 

- Did you add the side_channels argument in self._env as per the final bit in my original post? - In the init file.

 

From my perspective I added everything as suggested.

1. The import in the header
2. dimensions, timescale, and set_config in the __init__()
3. the channel = EngineConfigurationChannel()
4. the if (set_config == True)
5. The side_channels in self._env()

Could it be that I use the class wrongly in my notebook?

Following code copied from my file:

Code (Python):

1.  
2. class UnityEnv(gym.Env):
3.     """
4.     Provides Gym wrapper for Unity Learning Environments.
5.     Multi-agent environments use lists for object types, as done here:
6.     https://github.com/openai/multiagent-particle-envs
7.     """
8.  
9.     def __init__(
10.         self,
11.         environment_filename: str,
12.         dimensions: int = [],   #Added
13.         timescale: int = 1,     #Added
14.         worker_id: int = 0,
15.         use_visual: bool = False,
16.         uint8_visual: bool = False,
17.         multiagent: bool = False,
18.         flatten_branched: bool = False,
19.         no_graphics: bool = False,
20.         allow_multiple_visual_obs: bool = False,
21.         set_config: bool = True,    #Added
22.     ):
23.         """
24.         Environment initialization
25.         :param environment_filename: The UnityEnvironment path or file to be wrapped in the gym.
26.         :param worker_id: Worker number for environment.
27.         :param use_visual: Whether to use visual observation or vector observation.
28.         :param uint8_visual: Return visual observations as uint8 (0-255) matrices instead of float (0.0-1.0).
29.         :param multiagent: Whether to run in multi-agent mode (lists of obs, reward, done).
30.         :param flatten_branched: If True, turn branched discrete action spaces into a Discrete space rather than
31.             MultiDiscrete.
32.         :param no_graphics: Whether to run the Unity simulator in no-graphics mode
33.         :param allow_multiple_visual_obs: If True, return a list of visual observations instead of only one.
34.         """
35.         base_port = 5005
36.         if environment_filename is None:
37.             base_port = UnityEnvironment.DEFAULT_EDITOR_PORT
38.  
39.         channel = EngineConfigurationChannel()        # Added
40.  
41.  
42.         #Added
43.         if set_config == True:
44.             channel.set_configuration_parameters(time_scale=timescale, width=dimensions[0], height=dimensions[1])
45.         #Added
46.  
47.         self._env = UnityEnvironment(
48.             environment_filename,
49.             worker_id,
50.             base_port=base_port,
51.             no_graphics=no_graphics,
52.             side_channels=[channel],        # Added
53.         )
54.  

 

Ok, I found my foolish mistake. Of course, in my notebook I have to create my gym environment like this:

env = UnityEnv(env_name, worker_id=1, use_visual=False, multiagent=True, timescale = 1)

Guess my python is quite rusted.

Now I just have to figure out, how the timescale actually scales.