bzier
commented
6 years ago First I want to lay out a couple basics with the intent of each of these projects, the way they are currently set up, and how they depend on each other.
- TensorKart has several aspects:
record.py- used to record samples of human driving. Captures emulator screen images and associated human controller data.train.py- trains a neural network (supervised learning) using the recorded samples of human driving. Uses the images as inputs and the human controller data as expected outputs.play.py- loads a trained neural network model to drive races, feeding emulator screen images as input to the model, and the model produces controller output. This component depends on gym-mupen64plus.
- gym-mupen64plus is a wrapper for the mupen64plus emulator that provides an OpenAI gym interface to AI agents for interacting with the environment (in this case Mario Kart 64). It handles feeding the agent screen images as 'observations', and expects controller data as input, an 'action'. This depends on the mupen64plus-input-bot for communicating that controller data to the emulator.
- mupen64plus-input-bot is used for connecting the mupen64plus emulator to an HTTP server as a controller. Instead of a physical controller/gamepad, an HTTP web server will be used to serve up the controls.
Ok, so now onto your 3 questions.
-
I'm not sure what's going on here. First of all, the left terminal looks fine. Because you are attempting to use
record.pyto capture your own human driving, we would expect the emulator to be using the 'normal' SDL input plugin (not the mupen64plus-input-bot at this point). This means the emulator will be reading the USB controller gamepad, which is what we want. However, therecord.pycomponent is also going to be reading that USB controller (to record your human actions). That is where it is failing (in the right terminal). It looks like parts of the stack trace are cut off in the image. Are you able to copy/paste the whole error that is happening? -
The second error you reported (
ImportError: No module named 'BaseHTTPServer') is unfortunately a Python 2 -> 3 incompatibility. This is coming from the first line of the gym-mupen64plus project'smupen64plus_env.pyfile. It needs to be changed# From Python 2x: from BaseHTTPServer import BaseHTTPRequestHandler,HTTPServer # To Python 3x: from http.server import BaseHTTPRequestHandler,HTTPServerThere may be many more things like this that you will bump into with using Python 3. Again, if you are able to work through them, I'm open to a pull request. However, I'd like to maintain support for Python 2 as well. This can be accomplished with something like:
import sys if sys.version_info[0] >= 3: # Python 3 specific definitions ... else: # Python 2 specific definitions ... -
Yes, you are correct. When running
play.py, the terminal will show the AI model output in green, and human overrides in yellow. By the way, those aren't actually one-hot vectors, they are the control values themselves (e.g.[-75.3, 2.3, 1, 0, 0]). A one-hot vector would be a single 1 in a series of bit values (e.g.[0, 1, 0, 0, 0]). In the case of TensorKart, the first two values are the joystick x/y axes (continuous values between -80 and 80) and then three button for A, B, and RB (discrete values 0,1). I do have environment options within gym-mupen64plus for a discrete action space, which is where an agent chooses the best action among a discrete set, as opposed to directly choosing the controller values themselves. In this case, it actually is a one-hot vector, since it picks a single action from the set. Alternatively, this can just be represented as an integer value (the index of 1 in the one-hot vector).
tanakataiki
@bzier
I use Nintendo64 Controller and Following Error occurs
but I assume mupen64plus-input-bot should be used for 64 controller. I already set them up and I dont know why this happens. Anyway In my case R->Start, A->B and I cant press A in any other button and I can use joystick as usual, I use One controller though it suggest two controller is found.
I am new to gym so I want to know the way to use Test code I create this file in gym-mupen64 folder and I ran
but
ImportError: No module named 'BaseHTTPServer'Occurs.Would you tell me TensorKart Needs Gym Environment and should run code for gym or just need manual input data of input button and visual records ?
Sorry that I am a bit confused with TensorKart Instruction.
Thanks master !