If you haven’t already seen the video, make sure to watch it here!

Now You’re Thinking With Lasers

Its hard to convey the magnitude of this 1500W blue laser. To put it into perspective, this laser’s output power is almost exactly 2 horsepower. That’s about the combined power of 3-5 electric scooters at peak power. Continuously.

This laser can turn a fire brick into glass and punch holes in sheets of stainless steel. It does this by focusing all that power into a small area, allowing the surface of an object to reach ridiculously high temperatures if only you keep the laser pointed at an object long enough for it to heat up.

This is fine if the object doesn’t move, but I will need to give the laser a mind of its own if it is to handle moving objects. Working with the great people behind the youtube channel Tech Ingredients has given me the chance to do so, supplying me with the laser assembly and tasking me with intelligently controlling it.

“AI” Object Detection

Nowadays, everyone uses the word AI for everything that has anything to do with machine learning or neural networks. Believe it or not, when I took my first class in machine learning at MIT in 2019 they told us that using the word “AI” to describe machine learning was tacky. I guess that’s not true anymore since chatGPT. So I guess its called AI now.

This project uses an open source object detection model called “YOLO” which stands for “you only look once”. This name makes sense when you consider how the network works. As I explained in the video, YOLO works by taking an entire input image and making all detections for that image in a single pass. This allows our laser to make sense of what it sees.

YOLO’s Architecture

To learn how YOLO works, read this site.

The Brain of the Project

I got this RTX 3070 GPU from swapfest from someone who knocked off a capacitor on the board while trying to install a watercooling block. I got a good deal on it, and I knew I wanted to use it for some machine learning project in the future, although I did not know at the time what I would use it for. In the case of this project, it proved to be very useful in running some very powerful state of the art object detection and segmentation models. You really don’t need a GPU much better than this to get performance similar to top end GPU’s since object detection models don’t need that much vram to run inference.

Some Hardware

I also made this belt-tensioning 3d printed mounting mechanism for nema 23 stepper motors.

I also designed and soldered up a “motion control board” which basically was a teensy 3.2 that made the signals to drive the stepper motors along with a boost converter and some transistors to control the SSR which controlled the main laser. This let me control the main laser from my computer by sending commands using my own serial protocol between the computer and the teensy. The motion control board also had the job of enforcing joint angle limits to alleviate me from thinking about it when sending commands from the main tracking script.

Putting it all together

Here’s a demo of the round robin tracking logic, demonstrating the ability of the tracking algorithm I wrote to remember which object is which through time.

Destruction

Joseph Redmon

I can’t talk about YOLO without also talking about its creator, Joseph Redmon. Although his work as a PhD student at the University of Washington has forever changed the field of computer vision, Redmon decided in 2020 to step away from his project due to ethical concerns about the use of his model. It is clear from what was demonstrated in the youtube video that the capabilities of YOLO are not limited to detecting tumors in patient X-ray images.

YOLO is currently being used for mass surveillance and autonomous weapons technology. People like me are able to download and fairly easily wield the immense power of YOLO. It is worth reiterating that before this model, it was quite impractical to wrap a high frequency control loop around a camera based vision system and expect fast and robust object detection performance. Redmon can’t change this fact; pandora’s box was opened when he published his thesis and source code.

A Big Thank You to Tech Ingredients

Huge thanks to Alex and Eric for bringing me on to work on this project. They are very smart and driven people. If I had the time and was in the area again, I would like to revisit this project and scale it up another order of magnitude.