Vex AI questions

Although we may be jumping the gun a little, our team is very interested in VEX AI and have a few questions.
First off, I noticed that micro controllers are being legalized. How would something like an arduino or esp32 interact with the brain? Unless some new expansion hubs are introduced, I really don’t see a connection being possible.
Secondly, for the actually AI portion of the game, we were wondering if the competitiors would have to create models themselves using self prepared data or if we will receive a pretrained model already downloaded onto the microprocessor.
In addition to the above comment, I was wondering if (assuming if we must supply our own data and train a custom model) the process of creating an object detection algorithm would be completely text based, with users using popular libraries (tensorflow, pytorch, scikit-learn etc.) or if the learning process will be completely “block” based.

I am sure more details will come out over the coming days but as it stands today, there is plenty of scope for interfacing with the V5 brain via USB or RS485. VEX U teams do it all the time.
From the information that is available, it looks like you will get access to plenty of tools as part of the team registration.

2 Likes

I think there are a number of misconceptions in your question- maybe I can help clear them up.

1.) Artificial Intelligence (AI) is not limited in scope to Neural Networks.
Rebecca Reynaso has a great summary here: https://learn.g2.com/types-of-artificial-intelligence.
Examples of non-neural AI include Finite State Machines and Semantic Reasoning, such as Petri-graphs, etc.
2.) Pretrained models are largely useless for actual application- either bloated in order to meet your desired scope, or insufficient. I can see pretrained CNN’s for ball detection to be mildly useful.
3.) Datasets are far more interesting than models. In any deep learning project, gathering data is about 90% of the work.
4.) For your scope, I would definitely lean towards Keras Tensorflow (2)- still code based, but a massive simplification.

I wish you the best of luck!

3 Likes

Thanks for the reply! I understand that AI is not limited to neural networks, I was just listing object detection as an example of one of the common things teams will attempt to utilize (atleast so I assume). I personally have used Tensorflow, pytorch, and a little bit of Scikit-learn for personal projects so I have some familiarity there and I think those library’s would be great for students to get accustomed to because they are currently being used in industry.
I also totally agree that datasets are the most time consuming of the majority of AI-based projects, and I really am wondering if VEX will allow the user to supply their own data or if all of the necessary data will be supplied.
I’m really just hoping that vex gives more freedom than a simple drag and drop interface as has been done in the past…

Wow, you’ve used quite a few frameworks- that’s great!

When it comes to teaching basic machine learning, I prefer Scikit and Keras for students, where it should be of particular interest to differentiate between Tensorflow 1, 2, and Keras. I wouldn’t recommend TF1 to any student!

When it comes to data, it may be a good idea to build a common data repository that students can add to. That would allow students to have access to a far better dataset than the RECF can provide. (For example, images of balls would have various lighting conditions, angles, etc.)

I’m not quite sure what you mean by drag and drop interface- but it’s been a few years since I was a student.

You could easily run a simple CNN on an ESP-32. If you want a dedicated AI chip, here’s a rather comprehensive list:


and an overview of Neuromorphic chips:

2 Likes