LiCam, LiGrip O2 Pro and an AI agent in the software – Cody McColl, GreenValley International

0:00Michal Gula: We are already in the third day of the INTERGEO. how was the first two days like? What, what was your, like, top moment?

0:09Cody McColl: Ah. I mean, man, it's been crazy as always. GVI has always looked after our software latest releases, so we've had nonstop traffic. I don't think I've been able to take a breath this entire time. It's been a great show, a lot of new products. We're excited about our new products, and yeah, we're happy to be here as always, so.

0:25Michal Gula: What's your feeling about, because, I just spoke today to, to one person who actually counted all the SLAM scanning companies here at INTERGEO, and there were, like, 45, I think. So it's almost like a 10% of the whole INTERGEO being the SLAM scanners. How do you feel about that?

0:42Cody McColl: Yeah, I mean, the market is getting flooded, that's, that's very true. But I mean, GVI has always stand behind our self-developed algorithms. So our SLAM algorithms, we still feel are partened unto no one else, so I think that's what separates us, and plus the quality of our products mixed with the total ecosystem from GVI, I think we still have no fear.

1:00Michal Gula: look, the more competition there is, it usually means that there is, demand and there is a market for it, right?

1:07Cody McColl: Exactly.

1:08Michal Gula: So, like, your only job is to not to suck and then do your best job to deliver, right?

1:13Cody McColl: Exactly. And unlike a lot of the other hardware companies, because we have a post-processing software, we can still work with those products.

1:20Michal Gula: Perfect. So, what did you bring t- this year to INTERGEO?

1:23Cody McColl: Yeah. So we actually have several new releases, starting off with our LiCam, our spatial camera. This is just being released. Three cameras together. We have a LiDAR sensor on it as well that can actually be utilized. RTK GNSS capable, designed specifically for 3D Gaussian modeling, but it can be used at survey grade as well, plus minus two centimeters of accuracy. we've also-

1:42Michal Gula: At, at, at what range would you, like-

1:44Cody McColl: 40 meters.

1:45Michal Gula: 40 meters?

1:45Cody McColl: 40 meters, yeah.

1:46Michal Gula: but not all the data would be usable, like, up to 40 meters. Like-

1:49Cody McColl: Yeah

1:49Michal Gula: up to 12, 20, for example, or?

1:51Cody McColl: 20 is easy easy. Yeah, easy easy. I w- I would even say that 40 is reliable, but yeah, anything past that, the air will get substantially worse.

1:59Michal Gula: the RTK is not, included in the, in the device. Like, you have to mount it on it, right?

2:04Cody McColl: Exactly, yeah. On the backside, we have a USB-C port that'll allow you to attach an RTK device. 'cause a lot of, the design for the system was indoor RTK GNSS denied environments, so that's why we don't have it on the system now.

2:16Michal Gula: Okay. Ma- makes, makes sense. so y- you said that it's a spatial camera, camera, but maybe for the audience it's easier to say that it is for 3D Gaussian splats, right?

2:26Cody McColl: Absolutely.

2:26Michal Gula: For 3D Gaussian splatting.

2:27Cody McColl: Ex- exactly. With our latest advancements in LiDAR360 MLS version 10, the quality of 3D Gaussian is so much better that we wanted to create a device that specifically made that a true potential. So utilizing the true potential of our software was this device, as well as our next gen of Light, the LiGrip O2 series.

2:45Michal Gula: What would you see, like, who, who would be the ideal customer for, for this one?

2:49Cody McColl: I mean, originally we were thinking VFX, we were thinking, you know, camera, like gaming purposes. but the more that we talked to clients, we found that real estate, is one of those key factors. Being able to quickly walk through a building, create a 3D, 3DGS model that has high, high quality, that they can easily share with their clients, is where we think that this will probably do the best.

3:09Michal Gula: what's the price tag on this device?

3:13Cody McColl: Plus or minus 5,000, we'll say.

3:14Michal Gula: 5,000?

3:14Cody McColl: US dollars. Yeah, yeah. We'll j- we'll say right there.

3:17Michal Gula: Okay.

3:17Cody McColl: that may change as time goes on. Since it's official launch right now, we're not gonna tag a price that's exact at the moment.

3:23Michal Gula: I guess it's coming also with some subscription, right?

3:26Cody McColl: well, it comes free with the pre-processing software, and then the post-processing software would either be annual or perpetual based.

3:33Michal Gula: Can you process it,... I, I guess, of course, you can process it offline, but can you process it also, also on the cloud?

3:40Cody McColl: Yes. So the, the key is that at this, launch, we launched our LiCloud hub-based series. So we're now offering cloud-based processing, so it can either be processed on the cloud or local with inside of MLS software.

3:52Michal Gula: Okay, perfect. I, I, I know GreenValley for several years now, so I know that you are not only working with the lower tier scanners and entry-level scanners, but you have a, also, more sophisticated solutions. So what did you bring new from this level?

4:09Cody McColl: well, actually our O2 Pro. Happy to announce about our O2 Pro. So the key factors between this one and our previous version of the O2 is this one is offering a.5 centimeter relative accuracy, so we're talking sub-centimeter accuracy from this system. Higher grade IMU, but the key factor is three one-inch CMOS Sony sensors that allow for a much higher quality and crisp point cloud, as well as a 3D Gaussian.

4:31Michal Gula: How many me- megapixels and, what's the size of the sensor?

4:35Cody McColl: it's a total, I believe, of 30. It's a s- s- it's about 24, megapixel panorama, but around 30-plus, megapixels as the total. and this, each sensor is a one-inch Sony C- CMOS sensor.

4:48Michal Gula: Okay.

4:48Cody McColl: So much, much, much higher quality than our previous O2 series.

4:51Michal Gula: how would you compare the 3DGS output from this one and, from the, previous, 3D Gaussian splats camera?

5:00Cody McColl: Oh, tenfold different. the key with this one, because the one-inch sensor is so much bigger, it's much, much more capable of handling low light situations. So outdoor, our O2 never had a problem with the 3D Gaussian. Indoor it may have lacked before, but with these new three-inch, one-inch, three one-inch CMOS sensors, we're getting higher quality detail in low light scenarios, and fine tune writing is really what we are going for. We want to be able to see from five meters away a writing that's around one centimeter. Not possible with our previous version. With this version, it's totally possible.

5:32Michal Gula: What would be the, the, like, one single feature that you are really proud of with this one?

5:37Cody McColl: The accuracy

5:39Michal Gula: The accuracy?

5:39Cody McColl: The ac- the accuracy. The relative accuracy, because traditionally, any system from Hesai is around one centimeter minimum. with new algorithmic changes as well as a higher, higher accuracy IMU, producing a sub-centimeter accuracy point cloud has been the industry's request for at least five years. So-

5:57Michal Gula: But this is, like, a typical problem because we already mentioned that there are, like, 45 different, companies with the SLAMs. Typically, they are using one of the top three producers of LiDAR, so, on the hardware side they are basically more or less the same, right?

6:12Cody McColl: Yeah.

6:12Michal Gula: So the magic is happening somewhere else.

6:14Cody McColl: Exactly.

6:14Michal Gula: So can you explain that?

6:16Cody McColl: Yeah, so I would say that the magic is purely in the software. These are just the hammer. The nail is where you're really getting the tools out of it, you know? and the advancements in our algorithm for the 3DGS module has tenfold great compared to even version nine six months ago. a lot of that is because of the recent advancements in AI, d- in AI, as well as we've refined our SLAM algorithms specifically to make it a much higher quality, and I think that, again, truly is what sets GVI apart from the rest of us.

6:44Michal Gula: one part is collecting the data and processing. The, a totally different game is what can you do with the data? And, I'm pretty sure you can tell me something interesting because I have seen some videos, I have seen some, some data, so, what do you do on the use case side of the business?

7:04Cody McColl: Absolutely. So I mean, traditionally a 3DGS model was really just a viewer, right? Like, you have an ability to fly through.

7:11Michal Gula: Yeah.

7:11Cody McColl: Couldn't really do anything with it. with LiDAR360 MLS, we have the ability to edit, to segment, to classify Gaussian data. We can merge multiple Gaussian datas now together. We can aerial and terrestrial fusion Gaussian data.

7:23Michal Gula: How, how would, how would you merge them?

7:25Cody McColl: it's automatic. It's actually automatic based on pose. So we can take a, a pose base, 3DGS from aerial and pose base 3DGS from terrestrial and merge them together.

7:35Michal Gula: But do you need to use targets or is it, like, cloud to cloud?

7:38Cody McColl: It's cloud to cloud.

7:39Michal Gula: Okay.

7:39Cody McColl: It's cloud to cloud. It's algorithmic. Similar to, like, an ICP, but, slightly different in terms of algorithmic, at least speaking.

7:45Michal Gula: So-

7:46Cody McColl: So-

7:46Michal Gula: I, I guess the same would, go if you are doing the one site on a, like, a multiple scans, right? Like, like, on, on multiple looks.

7:53Cody McColl: Exactly. So similar to how we're able to do automatic registration with our, SLAM devices, inside of the software we can connect multiple different scans together and link them just as you would a traditional terrestrial scanner, but all inside of the same software.

8:07Michal Gula: I have seen it also on a, on a banner and on the, on the big display that you have these trees and, and forestry, like, you ha- you have a forestry use cases. I think also your software is somehow specialized for that on one part.

8:26Cody McColl: Absolutely. So GVI, again, we started as a software company, and actually specific in post-process- post-processing forest metric data. So from there we built LiDAR360 and LiDAR360 MLS, but we've never lost our focus on forestry being that key factor. one of the key things is obviously our company was able to help discover the tallest tr- second tallest tree in the world, species, second tallest tree species in the world, map it, create a 3DGS model of it, and it's leading to the potential of a national park being created specifically for that tree. yeah.

8:56Michal Gula: When collecting data for the forestry, can you again combine it with the LiDAR data from the above or-

9:01Cody McColl: Absolutely. Auto-

9:02Michal Gula: With the photogrammetry from the above?

9:03Cody McColl: E- exactly. So several different ways of fusing the data together. It's automatic. We can do it by three-point location as well. So you can have three-point location from your terrestrial and your aerial, merge them together that way. Obviously traditional methods of using ground con- ground control points as well. but we make it to where that in environments where you can't get down there that easy with control points, you can still merge the data together quite easily.

9:25Michal Gula: When I came to your booth, you said something about AI, implemented already in the software, so tell me about it because that got me interested.

9:34Cody McColl: Yeah, so our version 10 release, we're super excited about it. We have created our own LLM model built on our server that allows for true agentic AI driven by natural language prompting.

9:44Michal Gula: Okay.

9:44Cody McColl: So you can say, "I want you to segment this point cloud, give me all of the DBH, and give me a crown area and crown volume." One sentence is gonna generate that, take the model on an LLM, but process on your local computer. So it, it gives this ease of use, but takes it into that next level, which has also allowed us to now do automated tree speciation for several different types of tree species, both from aerial and terrestrial data.

10:07Michal Gula: But since you said that it's based on a LLM, on a language model, I'm just guessing that it's more about the easy of use and about the prompting of what needs to happen inside of the software, right?

10:19Cody McColl: Yes.

10:19Michal Gula: But, as long as we want in the future for the software and the AI to understand, what's being seen on the screen and 3D data, is it going to base more on a, like, visual language model or...

10:32Cody McColl: It's a mix of both. So it's actually intaking what the point cloud is already, all of the metadata related to the point cloud, as well as visual data, and then the original prompt. From the original prompt, it creates a series of questions. Those questions are dictating on whether or not we have a built-in skill in the software already, or if we need to reach out to open up models to get that skill. So then it can be automatically downloaded, grab the skill, and run that skill even if that skill doesn't exist in our software.

11:00Michal Gula: How do you deal with the typical bug of the... bug of the AI, which is hallucination? Like, if it's not sure what the answer should be, will it tell you? Like, "I'm not sure."

11:10Cody McColl: Yes.

11:10Michal Gula: Or will it just figure something out?

11:12Cody McColl: Nope, it actually generates an error question. It says that it feels that it's not correct.

11:17Michal Gula: Okay.

11:17Cody McColl: "Please review it yourself." Gives you the chance for manual review, and then it'll give you a step by step of what should be the next best case. If there's a true error that it discovers, it will find a way to resolve that error by using question-based prompt for you to, as the user, to manually check.

11:32Michal Gula: Yeah. Is the language model built on, some existing language model, or did you, like, develop your own?

11:37Cody McColl: We completely d- developed our own. That is completely de- developed by GVI, housed by GVI on our own servers.

11:44Michal Gula: and is this included in a standard version of the software with, with the license or do you need to buy it as an add-on?

11:51Cody McColl: It's actually included with our professional license or, or the annual license.

11:55Michal Gula: Okay.

11:55Cody McColl: it's not included in the f- the free pre-processing side of the, the device, but it is included in the professional version of our software.

12:02Michal Gula: what's like the hot topic right now in, in your company?

12:07Cody McColl: Obviously 3DGS is a ongoing topic with everybody. but really it's what's next, right? So the fact that we are able to bring a sensor that's a SLAM device down to a sub-centimeter accuracy, it... The real question is how much lower can we go? Can we get that down to a terrestrial-based scanner? If so, then let's target that market.

12:26Michal Gula: I honestly don't believe, like, in, in getting to the same accuracy with the SLAM compared to TLS, because, we can't, like, really, fake physics, okay? It, it, it comes with some limitations, I, I guess. And also, like, is it really necessary to go, like, under the centimeter accuracy? Because if somebody is buying a SLAM and he understands the technology, they are not expecting millimeter accuracy, right?

12:50Cody McColl: A- 100%. 100%. I think the key is, is that if we can bring SLAM technology to that realm of sub-centimeter or less, then that gives it a lower cost p- price point, which means more people are gonna be able to use it, more systems in the market, and less need for really the knowledge of how to use a terrestrial scanner, the overall linking of terrestrial scanners, and the time-consumingness of terrestrial scanners. The original reason why SLAM was created in the first place, right?

13:17Michal Gula: I usually have problems with, SLAM companies that they are not that good in, training and explaining the limitations. most of them are trying to sell, like, one push of a button solution, which it really is not. Once you get into bigger projects, and, like, as soon as you start using SLAM you will get to b- bigger projects, then you can get in a r- a lot of trouble. So how do you on your side try to support newcomers to the business and, and teach them with, with... how to work with the SLAM?

13:48Cody McColl: Well, one of those would be creating as much marketing material in terms of YouTube videos. So we have hundreds of YouTube tutorials. We try to make them very specific to each tool, and a very detailed analysis of that tool in our YouTube vid- tutorials. But another way that we've made it very easy for users is that AI assistant tool can also just be used like a similar ChatGPT that's been trained on our software. How do I create a DEM? What do I use to create a DEM? What's the best practices? And inside the software, it'll give you that information. But to take that a step further, we have a built-in model builder. That model builder allows you to create your own model just by dragging and dropping, verifying and running, so that you continuously use the same model over and over.

14:28Michal Gula: Okay, okay. have you tried... Because, like, when I'm processing the data with some software and I have some problem, I usually go to ChatGPT or, or Cl- Claude, right? Have you tried to do the- comparison and, and some benchmark of how your internal LLM co-, works compared to ChatGPT or Claude?

14:46Cody McColl: yes, but we can't release that information. But yes, yes. We have done major deep dives in how we compare to ChatGPT. the reality is, is that ChatGPT is a global train, right?

14:57Michal Gula: Yeah.

14:57Cody McColl: So it's intaking so much information, so much training data, and it may-

15:01Michal Gula: And mo- most probably also from different brands, different, scanners, right?

15:04Cody McColl: Exactly. Where with ours, we hyper-focused on the training of our system, obviously using publicly known knowledge as well of best case algorithms, best case equations and formulas. but we made sure that our software was able to take this LLM and take that concept of LLM in a software for lidar to the next level.

15:21Michal Gula: What's the challenge for the next year with, GreenValley?

15:25Cody McColl: Ooh. Making sure that anybody that tries to produce their own version of an agentic LLM, that we continuously surpass them at every point.

15:35Michal Gula: Ah, okay. thank you very much. If, if you have some last call to action or whatever message you wanna deliver to the audience, it's your place right now.

15:43Cody McColl: Okay. Yeah. I mean, honestly, GVI is extremely honored to be at INTERGEO. Every year, we've constantly seen the, the upgrades in, in INTERGEO. we've constantly seen the upgrades in the industry as well. As you mentioned, 45 plus different SLAM scanners at one event under 45 different company names. it's crazy. So making sure that you can weed out the ones that may be falsely, stating information, I wanna make sure that we as a company can provide that to customers, that we give you knowledge enough to know, like, okay, it's dangerous to choose certain, misleading information, I should say. So I wanna make sure that everyone knows that they can reach out to GreenValley and have verification of, is this system worth it? If so, do you have any proof of that? Do you have verification? And we can provide what we can, at least at a, at a, at a legal level.

16:31Michal Gula: Thank you for the talk and good luck with your product.

16:31Cody McColl: Thank you, Michael. Pleasure.