Learning about Theia from Radenso's DSP Engineer

[Vortex]

Easily one of my favorite parts of the show has been sitting down and chatting with Rob, Radenso's DSP/RF engineer. This guy is incredibly smart and so passionate about this stuff. It's been a blast to just listen to him and learn all about Theia, how it works, what it's capable of, etc.

We wanted to sit down together and chat a bit about Radenso's new detector on camera so here's a video nearly half an hour long of the two of us just geeking out about this stuff.

After SEMA, we'd like to do more videos together. He's excited to share more about this and I'm excited to learn more. We're not totally sure on the best level to go into this. He loves going into super technical detail, but for this video he kept things relatively simple and digestible for people.

We've also found it's super helpful to have visuals like the "swoosh" when detecting IO or showing you the difference between a frequency modulated BSM and a continuous wave radar gun. Similarly, in this video when he's talking about differentiating signals by looking at the modulation at the base of the carrier, that sounds super technical, but when you can just look at two different CW signals and visually see the differences yourself, it makes things instantly clear when you can see unique shape of the signals just on either side of the big single spike. It's tough to put visuals together in the middle of the show, so after the show he'll be able to put together some really great visual samples for us in future videos to help illustrate and explain this stuff.

If you have some additional questions as well as feedback on how we can best present this type of information, please let us know what you liked and what you think we could try that would make this even better for you.

 

[tempnexus]

Easily one of my favorite parts of the show has been sitting down and chatting with Rob, Radenso's DSP/RF engineer. This guy is incredibly smart and so passionate about this stuff. It's been a blast to just listen to him and learn all about Theia, how it works, what it's capable of, etc.

We wanted to sit down together and chat a bit about Radenso's new detector on camera so here's a video nearly half an hour long of the two of us just geeking out about this stuff.

After SEMA, we'd like to do more videos together. He's excited to share more about this and I'm excited to learn more. We're not totally sure on the best level to go into this. He loves going into super technical detail, but for this video he kept things relatively simple and digestible for people.

We've also found it's super helpful to have visuals like the "swoosh" when detecting IO or showing you the difference between a frequency modulated BSM and a continuous wave radar gun. Similarly, in this video when he's talking about differentiating signals by looking at the modulation at the base of the carrier, that sounds super technical, but when you can just look at two different CW signals and visually see the differences yourself, it makes things instantly clear when you can see unique shape of the signals just on either side of the big single spike. It's tough to put visuals together in the middle of the show, so after the show he'll be able to put together some really great visual samples for us in future videos to help illustrate and explain this stuff.

If you have some additional questions as well as feedback on how we can best present this type of information, please let us know what you liked and what you think we could try that would make this even better for you.

Love the info and the opinions, but being a devil's advocate; Is Radenso worried that sharing too much info and the methods might endanger their IP?

Also, can you ask Rob when his phase cloak tech will finally be released? I mean that's the PhaseView right? The ability to Phase out of view? Also, how does he feel about violating the Treaty of Algeron?

 

[dougr]

Honestly astounding. I can't wrap my head around what's possible with such signal detail. The dynamic range, what can be said? That was my gnawing question. Un-bee-leave-uh-bull.

 

[thefrog1394]

Interesting! I think this video should help start to clear up some of the more technical questions getting thrown around on other threads. The amount of data available in these captured signals is pretty wild.

And I really appreciate Radenso being open to sharing technical details like this and potentially open sourcing some components. What a far cry from the days of Escort banning members from their forums for talking about a bug.

 

[dougr]

Love the info and the opinions, but being a devil's advocate; Is Radenso worried that sharing too much info and the methods might endanger their IP?

As Rob mentioned, this isn't revolutionary information. DSP techniques are standard in the military/3 letter world. But it still doesn't discount how shockingly detailed it is.

His comment about the RD industry operating on a gentlemen's agreement to not spend on R&D since the gravy train of sales was humming anyway is the most telling reason we haven't seen advancement till now. The others know what they need to do and can go find DSP/AI engineers as well. They have to put in the work.

 

[oversteer325]

Really great video. It’s all starting to make more sense.
I love how passionate the Radenso team is.

 

[Heywood]

That really put more pieces into place.

Pretty excited to see how they take our questions and feedback to give us the nuts and bolts.

Finally a company of enthusiasts that really are excited to do this. They all seem as jacked about this as we are.

Great video Vortex.

 

[dougr]

Sorry for a 3rd post so quickly, but this fact just blows my mind:

The AI is classifying the platform hardware, not the signal. Read that again. The _hardware_. Police radar can start transmitting to mimic BSM, or go frequency hopping, it doesn't matter. The hardware transmitter is ID'ed. OMG.

So we now have to worry about agencies repurposing Stanley garage door openers as guns. Mind boggling.

 

[AllAboutJake]

When Kuni left Uniden I suspected my next radar detector would be a Radenso and this just confirms it. Well done and well engineered. I'm very impressed.

I'm also impressed about the talk about modular, upgradable, and maybe even open source possibilities for this product. I feel like Radenso can be more open aspects of the platform because the training data and training protocol for Rai is where the secret sauce is. A trained neural net can be deployed in the field without revealing all the hard work and data it took to train the model.

 

[alloy00]

Sorry for a 3rd post so quickly, but this fact just blows my mind:

The AI is classifying the platform hardware, not the signal. Read that again. The _hardware_. Police radar can start transmitting to mimic BSM, or go frequency hopping, it doesn't matter. The hardware transmitter is ID'ed. OMG.

Exactly. CRAZY stuff here.

 

[Vancity23]

The openness demonstrated by Radenso, sharing the details of this new unit, educating us on what it is about and how it works, and having multiple employees so active on RDF, is just awesome. It really feels like we are in this together with a common goal, compared to other RD companies where we are just a number to them...

 

[Heywood]

Hopefully people now can stop fixating on the frequency.
It’s tough to wrap your head around something different when for years it’s been told it’s that one way.

 

[Vancity23]

Ha ha, old habits die hard...what is that expression, Humans sometimes need to hear things up to 8 times before it really sinks deep into our brains...

 

[AllAboutJake]

A trained neural net can be deployed in the field without revealing all the hard work and data it took to train the model.

Side note: Radenso has the potential to become the Tesla Autopilot of radar detectors. Deploy enough of these in the field, collecting real world data and returning it to Radenso to train and improve the model.

It takes a lot of effort (and a lot of radar guns) to train the model initially, but Radenso could achieve critical mass with real world data provided by customers.

 

[G37X Jockey]

Reminds me of the difficulty learning a new language. "But in English we say it this way".
Forget the RD language we all know. This tech is beyond everything we thought we knew. We're going back to school guys.

 

[dougr]

collecting real world data and returning it to Radenso to train and improve the model.

It takes a lot of effort (and a lot of radar guns) to train the model initially, but Radenso could achieve critical mass with real world data provided by customers.

I was firmly of this thinking as the hints were coming out. But now I don't understand how it would work. Given serious real world RF pollution, it seems to me the AI needs to train with clean signals. Imagine marking a "signal" and having it sent in, either automated or manual. How would we even know what gun fired? And how would the AI know what exact signal to train on in the mishmash of the whole picture of polluted RF being captured?

I don't know. Just more questions about such incredible tech.

 

[dudeinnz]

Great video! Thanks for sharing @Vortex and Radenso

 

[doubledge]

Now this is the stuff! What a great video. Thank you @Vortex and all the Radenso folks. Keep them coming!

 

[InsipidMonkey]

Fantastic video! Rob did a great job explaining the concepts in an easy to understand manner, which is an impressive skill as an engineer! Really looking forward to the follow up video(s) and visuals.

There are a number of highly technical members here, and I'm sure many would be interested in listening to Rob and Sterling go into way more technical detail if they have the time to share!

 

[AllAboutJake]

collecting real world data and returning it to Radenso to train and improve the model.

I was firmly of this thinking as the hints were coming out. But now I don't understand how it would work. Given serious real world RF pollution, it seems to me the AI needs to train with clean signals

The real-world approach only works when you have a massive amount of data, and this product category might not fit the bill. Its only with a massive amount of data where you can train through the noise (and even bad training data) You could have a "false alarm" button next to the "mute" button, and collect exemplars of false alarms... but you'll get some percentage of bad data when people press the wrong button.

Another approach is to look into the past to find patterns. For example, when you see a signal that's of high probability, go back and train against the last few seconds of data to see if you can improve range and detection. This approach might be successful in finding quirks of reflection or signal propagation that otherwise wouldn't be detected by a human.

One of the questions was regarding whether or not a radar gun manufacturer could simply make BSM look-alike. Not a concern apparently, but there are adversarial attacks for neural nets. Much like fooling Tesla's autopilot by intentionally putting stickers on the road or using a sticker to make an AI think a banana is actually a toaster. Its counter intuitive. For example: a subtle change to the noise characteristics of a signal might not fool the AI, but a spike in an unexpected part of the nearby spectrum might. I don't think this ends the cat-and-mouse game necessarily. The question is who is a more determined force. There are also adversarial training methods to help harden the neural net to these sorts of attacks.