shannonclaude

Super cool! Have been a fan of this project for a while. I think this is super important for anyone who is trying to do ml inference on edge hardware, but also the browser too.

1. MagNav needs to be done continuously over a trajectory. Like GPS, it actually needs to be fused with INS data to work. Also GPS actually takes longer for a good lock on. What your phone is actually doing is sensor fusion (GPS + wifi, onboard accelerometer and gyro, etc.) and caching of the satellite information in order to produce a best estimate immediately. So when you first get a lock it takes a while even up to 15 minutes depending on the environment. Afterwards your phone/receiver tends to remember what satellites there were in the trajectory. It's a background process thats constantly running. We've experienced this first-hand in our experiments.

2. Since I work for a company in this area, I can't say too much, but SWaP is definitely going to decrease. Right now I think you could decrease size to a midsize cereal box. I don't think a super accurate version would exist for phones yet. Power consumption is somewhat low, but you need a serviceable GPU and RAM, like an Apple M3 to run things. Power consumption will improve with improvements in compute.

I think cost is the most prohibitive of them all, because quality magnetometers and INS are not cheap. > $20K per unit at least. i'm not talking MEMS grade stuff you see in phones and student projects, I'm referring to Tactical/Navigation grade—the stuff you see on military platforms. I think you'll see MagNav on military planes much faster, due to their mission-critical needs. For commericial planes to use them at scale, one needs FAA clearance which is a bottleneck for good and bad reasons.

We're not in a perfect world yet. MagNav will take some time to develop. What I think is missing is a larger community to work on this issue. There isn't that much motion in the space. Less than 5 serious players in the game, IMO. There also aren't that many suppliers who make the parts (magnetometers, etc.) you need for MagNav.

Lots of more thoughts here, but if anyone wants to chat more, contact me at sir.claude.shannon@gmail.com

They change over geological timescales—so not anytime soon. We're talking millions of years, w/ movement of tectonic plates and such. So for right now, once an area is mapped, its mapped for good. HOWEVER. Maps may need to be redrawn if the quality of them sucks.

Currently 20-40 pounds right now, but we're working to reduce the SWaP. MagNav does rely on having accurate magnetic anomaly maps, which the government and certain private companies have access to.

We're working on making MagNav a one stop shop backup for GPS. I think to cover ALL cases however, you'll need other technologies. TERCOM, visual, and celestial all have their niche use cases.

But for most cases, MagNav should do the job. Happy to answer more

It's the Air Force's higher ed institution. They offer masters and PhDs. They have their own labs doing R&D.

So the role that magnetometers play in sensor fusion w/ IMU data is for yaw/heading/magnetic north estimation. In short, it aids your orientation (RPY estimate). However, with MagNav, they play a large role in supplying information that allows you to decrease your drift rate.

https://www.sagemotion.com/blog/how-does-imu-sensor-fusion-w...

Magnetic-anomaly based navigation (MagNav) is a real thing that can solve this problem, and has been shown to work with the accuracy of a few hundred meters. Perhaps the government and defense contractors should look into this technology more. With a few more years of funded R&D and FAA-certification, I think its pretty likely that we'll see some of these systems on planes soon. The military is already flying with it during their exercises.

disclosure - I do work on a team developing MagNav, but much of the seminal research has come out of the Air Force Institution of Technology. They performed it on an F16, paper results shown here https://ieeexplore.ieee.org/document/9506809

+1 to the comments here. Part of the issue here is running these applications in Python. It's not really optimized to handle these loads and do DSP-based compute efficiently.

I mean why not just use ForScore on an IPad? It's the best solution available by far. I get this is probably a desktop app, so the comparison is not relevant.

Agreed—I got it just for that too!