Pro-tip - at this point take a step back and research which platforms have worked with mostly or fully autonomous control of robot from AI and VEXU programs - then figured computation complexity of their approach and ask yourself, will it be V5RC legal on existing platform.
I would look carefully at V5 Brain architecture - will it be good to handle the challenge you propose? My gut, a long reach right now.
Prove me wrong!
I am now attempting to copy a neural network code to try and understand how they even work.
Iâm trying to do it in C++ but I have a problem. I donât know how to code in C++ as Iâve never even used any kind of text code. Plus I donât understand how neural network inputs work in text coed so Ima just try and do it in blocks
I am beginning to barely understand what I am doing.
Thatâs probably not a good sign, you have to learn to crawl before you walk, and you have to know how to walk before you run. This challenge is not a walk, nor a crawl, this is more like being a cirque de soleil performer with one limb. The first step for this is probably codding a simple AI in python, if you donât know python, learn it. You are going to have to code this AI in a much more advanced language then block. Probably C or Lisp. You will need to pull a John Carmack.
Honestly good on you for trying this. I would recommend maybe having chatGPT explain some stuff to you. The code it gives you will be garbage, but usually it can point you in the right direction.
I understand what you are trying to say, however, I learned to code in the first place via trial by fire, and if I do say so myself it worked very well. So while yes I very much understand this is a very stupid thing to do for my first ever text project, I will be doing it anyways because I have nothing better to do in my free time & want to experiment.
Anybody know how I can get the weights & biases to save between runs?
Iâm copying off googleâs generative AI; how bad is that?
I agree with @Forgotten_Mail7702A on this one. When I first started programming, I wasted too much time trying to cheat my way out of learning. I hate to say it, but block coding is going to take you nowhere. I donât care if X team did Y thing with block, you should still learn C++/python. Itâs great for beginners, but once you want to do more advanced stuff (like neural networks) itâs necessary. I have seen too many people regret not switching to C++ sooner.
This is genuinely good advice. Make sure not to get stuck using AI generated code. Having AI generate all your code for you will leave you confused and frustrated. ChatGPT is not a programmer. Itâs a programming tool.
I say these things because Iâve made these same mistakes
As previously stated, I am using C++ for this code, however I also do not know how to use C++ and will learn how to via trial by fire because 1. I have nothing better to do in my free time. and 2. I am clinically insane.
I think a bigger concern would be how to get enough information about the surroundings to feed to the AIâyou donât have a lot of sensors to work with in VRC. If the AI canât determine whatâs happening, processing power isnât going to matter.
You really should youse python for machine learning as there are much more resorces than with C++
Im just using 3 AI image sensors, a GPS sensor, & 4 distance sensors. The AI sensor already has an image recognition Sofware build in so I donât need an AI for that part. Then the neural network and reinforcement learning part should hopefully figure out the rest.
Logan Sargent plus Lance Stroll = AI driven car.
The google generative AI wouldnât generate anything in python & C++ is better for performance which is important to get this too actually be able to run on the brain without it having the reaction time of an inanimate object.
Guys please how do I get the weights & biases to save inbetween running the code if i dont get it to save there will be no point in even using text code and I would have a better time individualy coding each neuron and adjusting all the weights & biases on my own and I dont have 2 years to do that help.
The best way would be to use an SD card and save it to a file for training. Once the AI is done training, you could embed the weights directly in the code.
You may want to look into training the AI virtually, and then transferring the weights to the robot, as training with the physical robot will be extremely time consuming.
Itâs been done in VRC 
Copilot: A ML library for Vex V5
This repository was used to control a robot in the World championships two consecutive years, in high level play resulting in a division Create award at Worlds. Feel free to reach out with any questions.
Also worth noting this repo is not in active development but could still be a good tool.
Yes, the AI doesnât think like us; it doesnât use strategies that we do, such as goal stealing into neg corner. It would be a better timekeeper and tell you how many goals, rings on goals, or rings on wallstakes and can maybe help you to keep track of time and say âT3 nowâ or smth. It wouldnât be bothered by pushing and shoving by other robots and probably wouldnât care about trapping. I think you would need a lot of sensors, and a lot of hard work and research, but it could be beneficial.