An alternative method is to have the robot run at random for n trials and examine the top 10-20 trials among those n trials. In doing this, we establish what is effectively a more complex version of AI Flappy Bird in expanding a population of generations that get increasingly successful from an initial generation that is completely random.
The core pillar of most autonomous robotics is figuring out how to use and translate sensor data. In our case and others, we have a relatively simple set of outputs - perhaps half a dozen motors. It's the whole process between the raw sensor input and raw motor output that makes an autonomous machine.
By combining standard V5 and the AI sensors, we have a fair assortment of sensing capabilities. For example, the AI Depth camera helps us keep tabs on field objects within a viewcone.
Though it's far from a concrete road, there are two general bounds to how much control the AI has over our robot. No control, and it defeats the point of the AI: all sensor data is analyzed by pre-programmed mathematics with no room for change. On the flipside, full AI control: both the raw sensor input and raw motor output are wired directly to the AI with no guiding hand of pre-written code.
Our general idea is a sort of middle ground, where we think of it as a control panel. In the first part, the AI is fed a very slightly refined sensor input (focused mainly on the VEX sensors, which are less refined and comprehensive than the new AI ones). In part 2, we feed the neural network a few "constant bias(es)", modifications to the weights of certain inputs that are out of the AI's control. A subtle set of constant biases would help with executive tactical decisions, allowing us to essentially give commands to our robot to follow a certain game plan while still leaving actual gameplay up to its discretion. And finally, we condense some of our outputs into commands - grouping motors of similar function together. For example, in a situation where we have a two-motor lift, we would instead give the robot control of a single "lift" motor which operates both - taking some load off the AI, making training simpler, and avoiding mechanical issues. This is also how we can implement the rules that Mr. Marshall discussed - with executive logic in the sensor, decision and action stages can tell an underdeveloped or troublesome network to do or not do certain things, like stop it from continually driving into a wall or overexerting its motors.
None of the above AI Control section covers anything about the actual decision making, but perhaps it will be useful to implement later.