Atari 2600 data collection for reinforcement training

Website atarigrandchallenge.com , designed to collect data

Hello! “Artificial Intelligence defeated a man in Go”, “Artificial Intelligence plays Atari 2600 better than a human being”, “Computer bots have approached human players in terms of the game of Doom” - the last time there are more and more such headlines. Numerous environments for developing and testing reinforcement learning algorithms appear (Reinforcement Learning): OpenAI Universe , Microsoft Minecraft Malmo , DeepMind SCII . And who knows what will happen tomorrow?

For my master's work, I want to collect a collection of replays of games for the Atari 2600, played by people. In the future, I will use it to train bots and lay out all the collected data in open access so that everyone can use them for their research.

For this purpose, the site created atarigrandchallenge.com . In fact, the site is a wrapper for the JavaScript-emulator Javatari, which for each frame of the game records all keystrokes, as well as points in the game, number of lives and other information about the game (for example, is the frame the last frame of the current game).
')
Based on this information, the sequence of frames of the original game will be restored. The bot will learn the game based only on the pixels of the screen and keystrokes.

If you want to help research, just go to the site and play a couple of games. To play you need a keyboard. The emulator does not work on mobile devices.

If you have any questions, ideas or criticism, write comments. Thank!