AI is pretty cool, right? Well, what would be even cooler is using it to detect different objects on the FTC field! That is what we plan to use it for.
We thought of the idea to use AI earlier this summer after the previous FTC season ended. We thought it could be helpful during the TeleOp and Autonomous periods, as long as it didn't become Ultron and try to destroy humanity and the other robots. Last week, at our second meeting, we took lots and lots of pictures of the blocks and balls to have enough data for our modified TensorFlow training script to run. By the end of the photoshoot, we had around 3,000 pictures, 1,500 of blocks and 1,500 of the balls. So I think we had enough data for the AI to train with. Next thing to do, install TensorFlow and start the training.
We decided to run a VM instance of Ubuntu 18, because Docker wasn't agreeing with our system. Once the setup of the instance was complete, we were ready to start the training. The training was the easy part. It only took around 10 seconds to fully train with an accuracy of 99.99%. We tested out some pre-shot images and the AI was able to identify them with very high accuracy.
This is a picture of the Terminal training window during the training process.
Next we plan on taking the trained AI and putting it into the FTC App Master to test with the phones camera and how we can implement the results from the camera into our autonomous and TeleOp. Check back in a few weeks to view how we are progressing throughout the season! Check our FaceBook and Instagram too!