Project Logs

Collapse

• Good News, Bad News (and an important update)

  Benjamin G. • 02/05/2026 at 08:30 • 1 comment

  It's been an eventful month for this project... after getting off vacation I ran into a bunch of problems, but made a bunch of solutions.

  
  TLDR: The prototype design sucks, so I made a simulator, adjusted part choice and configuration, added AI (the good kind). Details later. Please take the 1-minute survey.

  The Bad News

  • The prototype PCBs got lost in the mail. So I couldn't actually test the design.
  • Wouldn't have mattered much anyway, that design had a LOT of flaws.
  • Turns out, the current idea of using a radial magnet only provides 5DOF reliably. We will need to use a Halbach array of magnets instead to get full 6DOF. (5 magnets)
  • The ideal size for magnets for this design is 7x7x7mm cubes, which is not commonly produced.
  • The current choice of sensor is actually really noisy.
  • This probably needs 4 sensors, not just 3.
  • The usual method for estimating position and rotation from the sensors is too slow to run on a microcontroller.

  The Good News

  • PCBWay has graciously offered to send some free PCBs for prototyping, no strings attached.
  • I was able to build a simulator for optimizing & validating sensor and magnet designs.
  • Previous version didn't work well in simulation, due to aforementioned single-magnet 5DOF issue. I have created a design which works very well in simulations. More accurate. In theory.
  • We can get a MUCH better sensor for around the same price (TMAG5273). This has much less noise, more sensitivity, and includes a built-in averaging function. There's an even better version (TMAG5170) but it's 3x as expensive. We'll see if that's needed when I make the prototype.
  • Custom sizes of magnets can be purchased from Alibaba. The Halbach array with custom magnets is actually cheaper than the 2 radial magnets... in bulk.
  • It's actually pretty easy to train a neural net to be about as accurate as the conventional solver, and it's faster. I'll make another log with more detail later.
  • PCBA services are extremely cost-effective... in bulk. While it still is possible to create a hand-solderable version, there are a few pretty small/difficult components. It's very likely the version you will buy will only have the SuperMini board hand-soldered on. Maybe the buttons too. The rest will be professionally assembled at the PCB factory.
  • I am going to try and design this to be 100% 3D printed, no screws. Should be possible.

  I will create another update soon with more details on the simulator and neural net. Short version is, it uses MagPyLib and PyTorch to train a 36k parameter neural net to solve for position and rotation given readings from 4 sensors. It uses LiteRT (Tensorflow Lite) to run this on the NRF52840. The current setup is accurate to about 0.4mm and 2 degrees on average, without denoising.

  Here's a preview of the simulation visualization (using the neural net):

  Important: Take the 1 Minute Survey

  As I mentioned previously, this is actually pretty affordable, in bulk. Magnets for each board are about $1, NRF52840 board for $3, and the PCBA services can be less than $10. So we're still on track for the <$20 goal.
  

  HOWEVER, it's only cheap if I can order a bunch of them at once. Of course, this isn't the case for the prototypes, but once the design is finalized I will want to know how many of them to buy. Especially the magnets, as apparently they have a very long lead time (probably due to Chinese holiday soon). If I end up with 400 extra magnets, that'll kinda suck for me. If demand is stronger than I thought and I run out, that'll suck for you. As a college student, pulling the trigger to buy $100 in oddly sized magnets is not something I can do recklessly, so take the survey!

  Also, apparently the Magmouse name is taken, by a mouse with a magnet at the bottom, go figure... suggest names in the survey.

  https://forms.gle/YADcYZhiaJF6HaDy9

  Thank you.

View project log