Each tile corresponds to a 16x16 pixel square in the rendered screen, and since the top two rows (where the scores and other information are displayed) aren’t actually stored in the memory grid, we end up with 16x(13+2)*16 = 256x240 which is the pixel dimensions of the displayed screen. The actual displayed grid (16x13) scrolls through this grid once it reaches the end it wraps around and updates the memory grid (32x13) incrementally.

The tiles (blocks, items, pipes, etc) are stored in 0x0500-0x0069F as a 32x13 grid (technically it’s two 16x13 grids but it’s not difficult to get the correct coordinates).

I used the gym-super-mario-bros environment and implemented a custom observation method that reads data from the game’s RAM map. (Please refer to this issue for installing nes-py & pyglet.) Gym Environment smb-ram-ppo-train.ipynb.ħ.3.0 (7.4.0 was also tested and it worked) The pre-trained models are located under.

As of today (Aug 14 2022) the trained PPO agent completed World 1-1. My implementation of an RL model to play the NES Super Mario Bros using Stable-Baselines3 (SB3).

Super mario bros 3 world map module