Training Experimental Roadmap

[daniellawson9999]

Performed:

• 3 MuJoCo tasks
• Breakout (on a small, relatively arbitrary subset of steps converted to Minari)

What's next imediately:

• Convert more Atari data to Minari (The details of this and what we use, e..g all the data, filtered, what method we use to filter, have been discussed and should be elaborated on in a separate issue)
• Train and validate on a few subsets of games (we can do this locally). E.g; Pong, breakout, qbert, et.
• Move towards training on all 41 training games used by MGDT.
• Train on a combination of 41 MGDT plus several robotics tasks (TBD). We can use our 3 MuJoCo, probably want a few more. Our implementation is almost directly compatible with the official Minari implemented robotics environments*. We could do these official environments + MuJoCo we converted + our Atari suite

*Minari implemented robotics environments (https://minari.farama.org/) at this time: Door Hammer Relocate Pen Point Maze Kitchen