How Spatial Intelligence Is Transforming Machine Vision

WEBVTT
00:00:02.008 --> 00:00:07.087
Hello and welcome to Visions, a machine vision and automation solutions podcast.
00:00:07.621 --> 00:00:11.307
I'm your host, Jim Tatum, Senior Editor of Vision Systems Design,
00:00:11.580 --> 00:00:16.561
and Visions is an Endeavor Business Media production from your friends at Vision Systems Design.
00:00:17.205 --> 00:00:21.637
Here you'll find the latest on everything from end-user machine vision solutions
00:00:21.637 --> 00:00:26.673
to trends, developments, and perspectives on all things machine vision and imaging.
00:00:27.573 --> 00:00:30.663
Whether you've been working in the industry for a while or you're just starting
00:00:30.663 --> 00:00:34.873
to take a closer look at it, this podcast is designed to grow your knowledge
00:00:34.873 --> 00:00:38.962
and bring greater focus to your understanding of the imaging and machine vision industry.
00:00:39.630 --> 00:00:41.243
And now, on to our show.
00:00:42.108 --> 00:00:45.987
Well, hi everybody and welcome to Visions. Here's a question for you.
00:00:46.613 --> 00:00:49.852
What if machine vision could do more than detect objects?
00:00:50.491 --> 00:00:54.096
What if it could understand the world around it, predict what happens next,
00:00:54.520 --> 00:00:57.190
and help machines make smarter decisions in real time?
00:00:57.921 --> 00:01:01.056
Well, as it turns out, that's exactly what spatial intelligence,
00:01:01.178 --> 00:01:04.126
a rapidly emerging AI capability, does.
00:01:04.758 --> 00:01:09.003
And it's changing how vision systems, robots, and automation platforms interact
00:01:09.003 --> 00:01:09.890
with the physical world.
00:01:10.625 --> 00:01:14.247
Recently, Vision Systems Design's Head of Content, Sharon Spielman,
00:01:14.665 --> 00:01:18.770
got to spend a few minutes with Hantz Fevry, CEO of Geolava,
00:01:19.234 --> 00:01:23.210
a San Francisco-based company that specializes in spatial AI,
00:01:23.576 --> 00:01:27.802
for a first-hand look at how spatial intelligence is reshaping industrial imaging,
00:01:28.203 --> 00:01:31.870
system integration, and the future of AI on the factory floor.
00:01:32.446 --> 00:01:36.913
Hello, everyone. Today, we are focusing on spatial intelligence,
00:01:36.913 --> 00:01:42.829
the AI capability that lets vision systems understand and reason about the physical world in 3D.
00:01:43.676 --> 00:01:47.143
This breakthrough is influencing the future of industrial imaging,
00:01:47.143 --> 00:01:48.907
robotics, and automation.
00:01:49.296 --> 00:01:55.293
Joining us today is Hantz Fevry, the CEO of Geolava, to share his insights on
00:01:55.293 --> 00:01:58.839
how spatial intelligence is transforming vision system design,
00:01:59.019 --> 00:02:02.636
integration, and practical deployment on the factory floor.
00:02:03.048 --> 00:02:05.114
Thank you so much for being here today, Hantz.
00:02:05.915 --> 00:02:07.401
Thank you very much for having me.
00:02:08.233 --> 00:02:14.386
So I will just jump right in. In simple terms, what is spatial intelligence
00:02:14.386 --> 00:02:19.076
for machine vision, and why is it getting so much attention right now?
00:02:19.721 --> 00:02:24.346
Yeah, that's a good question. So spatial intelligence enables AI to perceive,
00:02:24.346 --> 00:02:29.066
understand, and predict the physical world across space and time.
00:02:29.495 --> 00:02:35.336
So it doesn't just recognize images. it represents the next evolution of AI
00:02:35.336 --> 00:02:41.597
moving from just the perception into meaning. So basically the reasons with the physical world.
00:02:41.963 --> 00:02:46.566
So if you think about AI systems today, you have one digital system talking
00:02:46.566 --> 00:02:48.457
to another digital system.
00:02:48.927 --> 00:02:52.671
Right now with spatial intelligence powered with world models,
00:02:53.031 --> 00:02:55.655
we're able to bring AI to the physical world.
00:02:56.235 --> 00:03:00.722
One key example is that a camera I can see a crack on the floor.
00:03:01.004 --> 00:03:05.856
A spatial intelligence platform powered by a world model will understand why
00:03:05.856 --> 00:03:12.682
it was formed, how it's evolving, and whether it will become a structural risk or not for a property.
00:03:13.431 --> 00:03:20.746
Okay. Thank you for that. So how does spatial intelligence enhance vision systems'
00:03:20.746 --> 00:03:26.318
ability to see and interpret complex industrial scenes? Yeah.
00:03:27.107 --> 00:03:31.156
So, of course, it gives machines the understanding of the state,
00:03:31.616 --> 00:03:37.853
the relationships, and the evolution of the physical environment that will allow them to basically,
00:03:38.476 --> 00:03:43.705
reason about the entire scene instead of just detecting objects in isolation.
00:03:44.186 --> 00:03:48.813
So instead of detecting, for example, a forklift and a worker separately,
00:03:49.092 --> 00:03:53.466
it will understand there is a collision path, and because of the collision path,
00:03:53.466 --> 00:03:54.455
there is something at rest.
00:03:54.740 --> 00:04:01.431
So there is a bit more of the reasoning beyond just the perception or the object detection. Okay.
00:04:02.496 --> 00:04:07.489
So what are the key technologies that are enabling the spatial understanding,
00:04:07.634 --> 00:04:09.288
especially in terms of vision?
00:04:10.339 --> 00:04:15.331
So it's a multi-modal platform. So a world model will focus a lot on the perception.
00:04:15.331 --> 00:04:18.621
So leveraging different types of imagery, it can be optical,
00:04:19.121 --> 00:04:23.891
LIDAR, thermal, 3D scene, and then as well as a contextual layer,
00:04:23.891 --> 00:04:29.191
it can also bring some text data and then basically understand not only what
00:04:29.191 --> 00:04:31.735
it's seeing, but the evolution of what it's able to see.
00:04:32.147 --> 00:04:35.841
So this is how you'll be able to build those spatial intelligence that just
00:04:35.841 --> 00:04:41.391
like a human, will be able to interact with the physical environment. Okay.
00:04:42.390 --> 00:04:47.699
So what have you seen to be the biggest practical challenges when integrating
00:04:47.699 --> 00:04:51.835
spatially intelligent vision systems in particular manufacturing?
00:04:52.451 --> 00:04:56.949
Yes. So the challenge will be, it won't be only collecting data.
00:04:57.269 --> 00:05:02.899
So it will be creating a reliable model that can continuously understand dynamic
00:05:02.899 --> 00:05:09.162
environment, that can adapt to change, and integrate with existing industrial workflows.
00:05:09.477 --> 00:05:12.808
So that most of the time are legacy systems.
00:05:13.191 --> 00:05:17.709
So that is a bit of a challenge where you have an industry that works with a
00:05:17.709 --> 00:05:20.819
lot of very solid workflows or systems.
00:05:21.184 --> 00:05:26.129
And then you have that frontier AI model that can perceive, that can reason,
00:05:26.129 --> 00:05:28.120
that can understand and predict.
00:05:28.561 --> 00:05:34.663
And how do you make sure that the legacy system can work in tandem with those new technologies?
00:05:35.290 --> 00:05:44.775
Okay. How does this spatial approach improve the flexibility and the accuracy over 2D vision systems?
00:05:45.425 --> 00:05:51.665
Yeah, it improves it tremendously. So traditional systems in 2D will recognize
00:05:51.920 --> 00:05:53.879
what's visible in a single image.
00:05:53.879 --> 00:05:59.199
So it will recognize an object. It will try to make some measurement depending
00:05:59.199 --> 00:06:00.593
on the type of imagery you're using.
00:06:01.011 --> 00:06:04.929
With a spatial intelligence, you can really understand the physical world itself,
00:06:04.929 --> 00:06:10.142
making it more robust, adaptable, and capable of anticipating future states.
00:06:10.485 --> 00:06:15.309
And basically what you can do is that you not only understand the relationship
00:06:15.309 --> 00:06:21.765
or the causal effects of a lot of features in the present, but also can forecast into the future.
00:06:21.962 --> 00:06:25.119
So you go way beyond detecting objects.
00:06:26.143 --> 00:06:32.035
All right. Can you talk a little bit about the role that AI-driven image processing algorithms play?
00:06:32.662 --> 00:06:41.438
Yeah. So, AI, they transform raw observations into a structured representation of the physical world.
00:06:41.879 --> 00:06:44.440
So, they can provide that foundation
00:06:44.860 --> 00:06:49.390
for reasoning, for forecasting or prediction, and decision-making.
00:06:49.657 --> 00:06:54.550
So with AI, you can be way more informed to understand not only what you see,
00:06:54.550 --> 00:06:59.392
but how it evolved into that current state and how it will continue to evolve
00:06:59.688 --> 00:07:02.087
into future states and forming better decisions.
00:07:02.721 --> 00:07:08.890
Okay. How can the spatial intelligence help vision systems better handle the
00:07:08.890 --> 00:07:14.029
tricky issues such as occlusions, reflections, inconsistent lighting?
00:07:14.673 --> 00:07:20.880
Yes so when you focus a lot one image or set of images then this is where you're
00:07:20.880 --> 00:07:25.080
gonna have those type of challenges but with a spatial intelligence or world
00:07:25.080 --> 00:07:27.720
model instead of relying on a single image you can,
00:07:28.380 --> 00:07:33.790
focus a lot on fusing information across sensors viewpoints time to infer and
00:07:33.790 --> 00:07:35.260
what basically cannot be,
00:07:35.740 --> 00:07:40.380
observed because the world model will be able to reconstruct beyond what it
00:07:40.380 --> 00:07:45.410
sees from the from the pixel so in that case those type of challenges are no
00:07:45.410 --> 00:07:48.260
longer present because of that reasoning aspect.
00:07:49.020 --> 00:07:53.900
Similar to a human being sometimes the human being doesn't have to see a place
00:07:53.900 --> 00:07:57.440
or a room with perfect lighting to be able to make an assessment you will be
00:07:57.440 --> 00:08:02.100
able to understand or even forecast what it should be what should be there,
00:08:02.580 --> 00:08:06.269
even without that perfect vision or clarity of the place.
00:08:07.002 --> 00:08:14.489
Okay, so it's not, well, maybe it is. It's using AI to replace what it thinks
00:08:14.489 --> 00:08:18.134
the lighting should be or replaces what the reflections are?
00:08:18.598 --> 00:08:22.529
No, so basically what will happen typically on world model is that if the lighting
00:08:22.529 --> 00:08:26.661
is not perfect there, so think of it as an image and there is a lot of mask
00:08:26.917 --> 00:08:31.709
on top of that image, it will be able to predict what should be there even without
00:08:31.709 --> 00:08:33.123
being able to see it perfectly.
00:08:33.419 --> 00:08:36.885
It's because you don't understand causality and relationships.
00:08:37.186 --> 00:08:41.092
So even if there's not a perfect environment, perfect lighting,
00:08:41.366 --> 00:08:45.127
then with that spatial intelligence, it can still navigate and see through it.
00:08:45.453 --> 00:08:50.099
But beyond that, the spatial intelligence, what it focuses on and what it's
00:08:50.099 --> 00:08:55.328
good at is basically it's on the reasoning. So it depends on what you're trying to accomplish.
00:08:55.705 --> 00:09:00.425
If you're trying to make an observation of a place to detect a risk of a structure,
00:09:00.742 --> 00:09:04.689
I don't know, any structural risk in that case, without the perfect lighting,
00:09:05.003 --> 00:09:09.495
you can still have a level of understanding of that risk instead of focusing on,
00:09:10.023 --> 00:09:12.895
a single image that will give you a binary answer.
00:09:13.639 --> 00:09:16.460
Okay. What about on the factory floor?
00:09:17.406 --> 00:09:22.602
More of our readers are into industrial manufacturing, so they're looking for
00:09:22.822 --> 00:09:26.629
a solution on the actual factory floor, you know, with the machine vision that
00:09:26.629 --> 00:09:31.367
they're using in, say, their inspection things on the manufacturing line.
00:09:32.654 --> 00:09:37.485
So if I understand the question correctly, it's what is the application for
00:09:37.485 --> 00:09:42.225
someone in the factory floor leveraging that compared to legacy system or legacy
00:09:42.225 --> 00:09:43.358
computer vision systems.
00:09:43.944 --> 00:09:48.279
So I would say right now in the factory floor, you can have a true co-pilot,
00:09:48.737 --> 00:09:53.745
that not only help you understand and make faster analysis, but make analysis
00:09:53.745 --> 00:09:57.245
for you and then help inform your final decisions.
00:09:57.542 --> 00:10:00.234
So then the person in the factory floor becomes way more efficient.
00:10:00.717 --> 00:10:05.415
And that spatial intelligence can be powered powering either a robot,
00:10:05.415 --> 00:10:09.645
so this is where we're talking a lot about physical AI, or it can be powered just,
00:10:10.285 --> 00:10:15.827
powering just a software and help make an analysis and help make some forecasting decisions.
00:10:16.053 --> 00:10:21.555
So this is some of the many applications that can be used when we bring world
00:10:21.555 --> 00:10:24.279
models to the factory floor. Okay, great.
00:10:25.364 --> 00:10:32.045
So what advice would you give engineers and integrators for validating and benchmarking
00:10:32.045 --> 00:10:36.231
these advanced vision systems? Yes.
00:10:36.765 --> 00:10:40.435
So this is where hallucination will come in. And I think what's very important
00:10:40.435 --> 00:10:44.788
is to benchmark real-world understanding, not just detection accuracy.
00:10:45.119 --> 00:10:48.665
So it's important to measure how well the system generalizes,
00:10:48.665 --> 00:10:53.382
reason about change, and predicts outcomes in an operational environment.
00:10:53.823 --> 00:10:58.265
And then always measure against those key benchmark, which is not easy,
00:10:58.265 --> 00:11:03.555
but also not impossible to do. So the real test is not whether it detects a
00:11:03.555 --> 00:11:08.584
defect today, but whether it predicts the defect before it becomes a failure.
00:11:08.950 --> 00:11:13.013
And what is that benchmark? What is that ground truth of a defect?
00:11:13.350 --> 00:11:14.654
So that will be basically...
00:11:15.867 --> 00:11:20.404
The advice I would share for any engineer who are trying to implement any world
00:11:20.404 --> 00:11:22.663
model or spatial intelligence system to their workflow.
00:11:23.490 --> 00:11:28.424
So from your perspective, how will spatial intelligence change the future of
00:11:28.424 --> 00:11:31.292
vision hardware and software design?
00:11:31.843 --> 00:11:38.494
Yes. So right now I think vision system or passive perception that detect objects
00:11:38.494 --> 00:11:41.282
that basically make sometimes some connections.
00:11:41.723 --> 00:11:47.014
But right now, those vision systems will evolve from passive perception tools
00:11:47.014 --> 00:11:52.654
into world models that continuously perceive, understand, and simulate the physical world.
00:11:53.159 --> 00:11:57.571
So just as language models became the intelligence layer for text,
00:11:57.780 --> 00:12:02.974
world models will become the intelligence layer for the physical environments. Okay.
00:12:03.706 --> 00:12:09.254
So what should manufacturing leaders focus on to build confidence and readiness
00:12:09.254 --> 00:12:13.759
for adopting these new spatially intelligent vision technologies?
00:12:14.714 --> 00:12:15.760
That's a great question.
00:12:16.942 --> 00:12:22.834
I would say to start with measurable business outcomes. To invest in high-quality
00:12:22.834 --> 00:12:28.774
multimodal data and prioritize platforms that continuously learn and improve
00:12:28.774 --> 00:12:31.003
as the physical world changes.
00:12:31.369 --> 00:12:36.714
So for example, so beginning with a production line where reducing downtime
00:12:36.714 --> 00:12:42.474
by 10% delivers an immediate and measurable return on investment rather than
00:12:42.474 --> 00:12:44.746
just testing the spatial intelligence,
00:12:45.300 --> 00:12:47.494
for the sake of having a new technology there.
00:12:47.494 --> 00:12:52.994
Because sometimes some legacy system are robust and work, and the spatial intelligence
00:12:52.994 --> 00:12:57.204
on a new system can only be useful in a very, very specific use cases.
00:12:57.204 --> 00:13:02.486
So basically identify those use cases and implement the spatial intelligence or world model there.
00:13:04.284 --> 00:13:09.755
Do you want to tell me a little bit about GeoLava and how long it's been around,
00:13:09.755 --> 00:13:11.015
how long you've been the CEO?
00:13:11.915 --> 00:13:16.225
You could tell us about that. Sure. So I've been working in AI for the past
00:13:16.225 --> 00:13:21.722
10 years. I was working at Google as part of Speech that is now part of Google DeepMind.
00:13:22.198 --> 00:13:27.535
So I work a lot on deep learning. So after Google, I built a company that we
00:13:27.535 --> 00:13:32.335
sold in 2024, and now we're building GeoLava. And DualLava, what we're doing
00:13:32.335 --> 00:13:35.173
is we're bringing AI to the physical world.
00:13:35.503 --> 00:13:41.505
We're building a world model for critical infrastructure and properties so people
00:13:41.505 --> 00:13:45.788
can not only understand their current state, but forecast their future state.
00:13:46.339 --> 00:13:50.972
So that's what we've been working on. And our team more from DeepMind,
00:13:51.192 --> 00:13:54.465
Apple, Meta, with very strong experience in computer vision,
00:13:54.465 --> 00:13:58.063
as you mentioned, and as well, AI in those frontier models.
00:13:58.724 --> 00:14:02.735
Okay, great. Right. Is there anything else you want to tell our audience?
00:14:03.115 --> 00:14:09.800
They are the engineers and the engineering managers at OEMs with machine vision and image processing.
00:14:10.357 --> 00:14:12.029
So what would you like to say to them?
00:14:12.692 --> 00:14:17.185
I would like to say to definitely make a lot of research on world models and
00:14:17.185 --> 00:14:21.625
not to confuse VLMs or LLM for world model. They're very, very different.
00:14:22.049 --> 00:14:27.635
And yes, as I mentioned before, to try to see where a world model or a spatial
00:14:27.635 --> 00:14:31.574
intelligence can bring value in measurable ways.
00:14:32.004 --> 00:14:37.945
And yeah, in my opinion, the world model will revolutionize the AI in the physical
00:14:37.945 --> 00:14:43.375
world the same way LLM revolutionized AI with X. Great.
00:14:44.002 --> 00:14:49.275
Well, thank you so much, Hantz, for joining us today and providing this information.
00:14:49.275 --> 00:14:54.295
It's going to be a really exciting time in our industry as we move into this
00:14:54.295 --> 00:14:58.873
physical AI space. So I really appreciate you speaking with us today.
00:14:59.490 --> 00:15:01.206
Thank you very much. I appreciate the time.
00:15:03.168 --> 00:15:07.249
Well, that's a wrap for this episode of Visions, produced by Endeavor Business
00:15:07.249 --> 00:15:09.931
Media, a division of Endeavor B2B.
00:15:10.273 --> 00:15:13.269
Thanks very much for tuning in. If you enjoyed today's show,
00:15:13.269 --> 00:15:17.079
be sure to subscribe to the podcast and share this episode with a colleague
00:15:17.079 --> 00:15:18.017
who would find it helpful.
00:15:18.679 --> 00:15:24.579
Until our next episode, you can find us at vision-systems.com or on LinkedIn,
00:15:24.579 --> 00:15:29.215
Facebook, or X for more insights, updates, and breaking news to keep you in the know.
00:15:29.987 --> 00:15:33.673
Thanks for tuning in. Until next time, stay focused on your visions.