what does the obs that is returned from step() function mean?

[ProQianXiao]

I am trying the Single Agent Example from https://marlo.readthedocs.io/en/latest/usage/singleagent_example.html, I run the example and I print the obs that is returned from step function like this. while not done: _action = env.action_space.sample() obs, reward, done, info = env.step(_action) print(obs) Then I get the output of obs, it looks like ndarray,

But what does that mean? The pixel point information of next state? Or others? Could you help me to settle that?

[spMohanty]

@ProQianXiao : This refers to the RGB frame representing what you would see on the screen of the game if you were playing the game yourself. Does this answer your question ?

[ProQianXiao]

@spMohanty Thank you for your reply. Can I understand this? The obs that is returned from step function represents the state information, and I could use it as a parameter to pass to my algorithm. Is it right?

In addition, I also want to know how can we get the detailed information about the findThegoal environment, such as description, actions available, rewards, discrete action or continuous action? I search the envs, but there is no findThegoal environment. Looking forward to your reply.

[douglasrizzo]

obs is an image, a screen capture of the game. Each of the three values in the lines from your terminal represent the R, G and B values of a pixel from the screen. If you do something like:

obs, reward, done, info = env.step(_action)    
im = PIL.Image.fromarray(obs)
im.save('image.png')

you'll see that a screen capture of the game will be saved as a PNG file. I hope that helps.

[ProQianXiao]

@douglasrizzo Thanks very much, that helps me a lot.

[douglasrizzo]

@ProQianXiao The FindTheGoal environment is named Basic in the screenshot you provided. However, I don't think you'll find any relevant info in the repository. You just have to interface with the game through the marlo package, send actions and receive rewards and new observations. I suggest you take a look at the usage examples. Personally, I don't think you'll have much structured data to work with (e.g. a discretized grid world). You'll just have to work with the image inputs.

[spMohanty]

Thanks @douglasrizzo for the clarification. And thats correct, in this competition, we are focussing on just image frames as observation.

At a later point in time, we do have plans to make the grid word available as observation (and make it configurable), but that will anyway not be allowed in this competition.

[ProQianXiao]

@douglasrizzo @spMohanty that is very kind of you. I figure out what the obs means and the how to use state information now. Thanks very much again.