Keeping the stack low is the easiest task in the game. Tetris scoring is designed to reward the risks involved in building a high stack.
You can't give 100% repeatable inputs as that opens the door to a fixed, hand-optimized placement pattern. If this is an issue, sample 100 runs and take one from the best, worst, and median cases.
What could be improved is to add some amount of "next piece" data. More data = more ability to use planning instead of probabalistic AI.
Initial configurations with garbage filling the bottom are a good way to test AI robustness, so I would recommend adding a mode for that.
Last, the most prestigious test would be to include the two-player modes. Like C-Robots or Corewars, but with Tetris :)