Using ML for Enemy Generation in Video Games

I am attempting to make a 2-D platformer game where the player traverses through an evil factory that is producing killer robots. The robots spawn at multiple specific locations in each level and impede the player's progress.

Enemies are procedurally generated using machine learning. Early levels have "garbage" robots that plop down and can't really do anything. After generations of training, the robots begin having more refined bodies and are able to move about and attack the player. Later levels produce enemies that are more challenging.

Enemies consist of a body and up to 4 limbs. The body is simply a circle of a certain radius, while the limbs are just a bar with a certain length. Limbs can pivot and/or contract/extend. Additionally, each limb can have one of three types of "motor" (wheel, spring, or hover). This makes for about 20-25 input parameters:

BodySize, Limb1Enabled, Limb1PivotPoint, Limb1Length, Limb1Angle, Limb1MotorType, Limb1MotorStrength, Limb2Enabled, Limb2PivotPoint, Limb2Length, Limb2Angle, Limb2MotorType, Limb2MotorStrength, Limb3Enabled, Limb3PivotPoint, Limb3Length, Limb3Angle, Limb3MotorType, Limb3MotorStrength, Limb4Enabled, Limb4PivotPoint, Limb4Length, Limb4Angle, Limb4MotorType, Limb4MotorStrength

My thoughts are that a genetic algorithm (or something similar) would be used to generate a body, while a neural network would control that body by using the same inputs to generate outputs that control the limbs and motors.

There would actually be 3 "control brains" that would have to be trained using the same inputs, but having different fitness goals: Moving Right/Left, Moving Up, and Attacking the Player. (Gravity exists in 2-D platformers, so moving down isn't necessary.)

A fourth, "master brain" would take the player's relative location, score, and maybe time elapsed, as inputs, and would output one of the goals for the robot to achieve (move left, move right, and attack).

The master brain's fitness would be determined by the "inverse" of the player's "progress", while each control brain's fitness would be determined by how well it was able to perform the task assigned by the master brain. Finally, the overall fitness for the body's genetic algorithm would be an average (or some other function like min, max, etc.) of the three control brain's fitness values.

Now that I have all this "down on paper", where do I start? I had planned on doing this in Unity, but early attempts have been a bit confusing for me. I've been able to procedurally generate a body with random limbs (no motors) that wiggle about randomly, but there's no neural network or any machine learning going on whatsoever. I am not exactly sure how to expose my parameters to be used as inputs, and am barely grasping how I should take those outputs to control what I want them to. Are there any libraries I should look at, or should I write this all from scratch?

Also, before I get too far ahead of myself, what are the flaws in my approach (as I'm sure there are plenty). I want my project to be something practical in scope, if training can't be done feasibly while a player traverses a level, this might just be a dead project idea.

Anyways, that all being said, thank you for your help.