Learning Agents - Cube to Cross a Narrow Path

It’s funny you created this gym because I have considered doing something similar at some point. You need to do the following:

  • Create the interactor BP - design the observation and actions that make sense (you should make your simple game playable by yourself as well)
    • Maybe do some downward angled raycasts to detect where the floor and the open areas are
  • Create the training environment BP - design the reward function
    • Reaching the goal is +1
    • Falling is -1
    • Timeout is -1

You really just need to copy the components of the racing environment and adjust the details to match, but overall the structure is the same.

Brendan

1 Like