bzier
commented
6 years ago Thank you for sharing that screen capture. That's incredibly helpful. I've seen symptoms of this before, but never actually seen what's going on to understand it.
The first thing to be aware of is how the gym environment is detecting the end of an episode (end of a race). At the end of a race, the game 'zooms out' to show replays and timings, etc. I am currently using this to detect when the race is over. See here at the bullet point for _evaluate_end_state(). That info is actually slightly out of date. It mentions all black pixels, but the code currently looks for all corners of the same color. See here for why.
Now at this point you may guess what the problem is. In your screen capture, you can see that the emulator is rendering briefly in full size and then flickering to a really tiny image, then back to full size, etc. The problem comes in when we encounter 30 consecutive frames of that 'zoomed out' tiny view and consider it the end of the race.
Unfortunately I'm not sure that this is a trivial issue as it is likely an issue in the underlying N64 emulator or video driver. Even if that behavior didn't trigger my environment's end episode detection, you likely would not get good behavior from any AI executing with that as the input... even a human would probably struggle significantly :)
I'll have to take a closer look at this later tonight to see if I can track down anything. Just be aware that it may not be a quick fix.
fujiki-1emon
Greetings,
I wanted to see some demo running and how the system runs and tried to run
python test.pywithout any inference from a model (fixing joystick = [0, 0, 1, 0, 0]). At first, it seemed well, but finally I got the problem below without seeing the agent moving around. It must be a trivial issue, but I'd like you to help me solve the problem.Thanks.