Automation as a Key Enabler ft. Juergen Daunis of understand.ai

[00:00:00] My name is Carl Anthony and I work in the automotive industry in Detroit. Sometimes that work encompasses future vehicle technology and that's what we talk about here for the most part anyway. This is AutoVision News Radio.

[00:01:00] I'm not a fan of AutoVision, but luckily I had a gracious guest. By way of introduction, Yurgen Donnis is the CEO of Understand AI.

[00:01:09] With domain expertise and automotive IT and telecommunications, Yurgen is focused on the crossing point between technology and business and has dedicated a significant portion of his career to connect it automated and shared services.

[00:01:25] As the CEO of Understand AI, Yurgen leads a world-class team that applies artificial intelligence and machine learning technology to change the landscape of the autonomous driving and ADAS annotation business.

[00:01:39] As described by Understand AI, in order for autonomous driving to be realized, an infinite number of images and point cloud frames must be annotated and must be annotated in a way that is fast, affordable at scale and with consistent quality.

[00:01:57] In order to train and validate perception stack and driver assist functions, founded in 2017 as a ground truth solutions provider, Understand AI is fueled by a vision that autonomous mobility can help support a safer and more sustainable environment for all road users.

[00:02:17] Moving at speed of mobility, this is AutoVision News Radio with Carl Anthony in Detroit, Michigan.

[00:02:23] Understand AI is founded in 2017 with the goal to make AI applicable in real road cases.

[00:02:31] We quickly focused on automotive and annotation business, especially in the fields of ADAS and perception development and validation.

[00:02:41] We produced ground truth data. The idea is that we can replace the majority of the manual work with AI and machine learning technologies.

[00:02:51] In most ADAS programs, the labeling today is still done manually to develop a perception stack. Usually you compare the device under test at the vehicle with reference system which produces the ground truth data.

[00:03:06] That's what we do. We produce ground truth data where we label cars, pedestrians, etc.

[00:03:12] Then you compare this with ground truth with the device under test.

[00:03:17] In our core product today, we have a very powerful offering where we focus on three aspects.

[00:03:23] So functional capabilities like multi-sense of fusion programs and this enables us to execute complex ADAS programs.

[00:03:32] The second point is operational excellence on tools, infrastructure processes, etc.

[00:03:39] And that helps us to be cost competitive and have a high data quality.

[00:03:44] And the third aspect is the automation. And that also has a big impact on the cost quality and the time to execute the program.

[00:03:52] Understanding AI has great competence around machine learning and artificial intelligence on cloud technologies, on software development, ADA, sensor domain expertise where we have also issued a lot of patterns.

[00:04:07] And we have recently expanded our focus also inside our core industry out of the motive to other industries like trains and heavy machines, agriculture mining logistics.

[00:04:17] And also what we do in our offering is pretty much comparable or adjustable for these kind of industries.

[00:04:24] On July 16, 2019, DSpace announced its acquisition of understand AI.

[00:04:31] The press release from DSpace at the time spoke about how understand AI would invest in artificial intelligence applications and cloud-based tools and further develop its existing products as an integral part of the DSpace product range.

[00:04:47] That same press release from DSpace explains that when developing an autonomous vehicle, it's crucial to detect the car's environment realistically and without faults.

[00:04:59] Likewise, other road users, traffic signs and static roadside structures and even open spaces must be reliably and accurately identified.

[00:05:10] As Jürgen explained, data-driven development was and still is a key aspect of DSpace's acquisition of understand AI.

[00:05:20] And data-driven development actually means that for example, for ADA, you need data to refine your models.

[00:05:27] It's not like linear code development. You train your models with data to then reach what you want to do.

[00:05:35] And to do that, DSpace has a couple of portfolio items and it starts with the data acquisition where there is a rooftop box with a logging device in the rooftop box where then you record the data in the vehicle.

[00:05:49] Then you ingest the data into the storage of the cloud, then you store the data.

[00:05:54] You have the data management and the selection of the right data and storage.

[00:05:59] And then you label the data, that's actually what understand the eyes doing.

[00:06:03] And then you can create the scenarios from real world data that you can then use in later steps in the process like in data processing part where you can use state of replay to navigate space testing simulation,

[00:06:46] understand AI's media packet contains a series of slides dedicated to the challenges the automotive industry faces regarding ADAS and automated driving functions that require validation through high volumes of annotated data, namely the challenges of time, quality and cost.

[00:07:06] So we see a lot of problems actually in the delay of special data acquisition.

[00:07:12] So to get the data in the right cost and the right timing and the right quality is quite complex.

[00:07:20] So this leads then, of course, to higher cost in data acquisition and the storage, etc.

[00:07:25] And there is a high volume of data actually that is needed to do these programs.

[00:07:30] But the industry challenge actually is make ADA does affordable, so get the right data or the right quality, right time in the budget, seek to the budget.

[00:07:40] The second field of complexity is the validation verification, home location and there's an ongoing discussion actually on how much real world data is needed to do that.

[00:07:51] But also can be used synthetic data, what percentage can we use it or can be used generative AI created data to do a home location or car and who is taking liability for that, etc.

[00:08:03] So this are open questions in the field and of course to do all of that only with real term real data is actually quite a cost driving.

[00:08:15] So it's not trivial to seem as the orchestra to fool data driven development in an efficient way.

[00:08:21] So for example, if you're developing a dust drive function like a distance control usually start with the ODD the operational design domain.

[00:08:30] Then you define the function and then you go downstream actually to find a campaign.

[00:08:36] And you go up the stream again, you record the data and then you load it up and you store it and manage it and select the right error and then label it and so on.

[00:08:46] Right, it's you need to think this from the end from the definition of the ODD and the function that want to reach and then you need to populate this from the beginning in the chain and to do this in a seamless and cost efficient way that is not trivial.

[00:09:00] During my time with Jürgen I was introduced to something called a CICD tool chain.

[00:09:06] If you are new to this concept or want to brush up on your knowledge, I will leave a link in the show notes to an info world article written by author Isaac Sikolik.

[00:09:16] I'm quoting Isaac's article now.

[00:09:19] Continuous integration, CI and continuous delivery, CD also known as CICD embodies a culture and set of operating principles and practices that application development teams use to deliver code changes both more frequently and more reliably.

[00:09:38] Continuing from Isaac's article and I quote CICD is a best practice for dev opt teams. It is also a best practice in agile methodology by automating code integration and delivery CICD let software development teams focus on meeting business requirements while ensuring that software is high in quality and secure.

[00:10:02] In the context of the vehicle and software defined car it's actually two industries coming together on the software development industry and this is also of course heavily also promoted by companies like Tesla.

[00:10:14] And then a more traditional automotive company where the V model and more strict development process.

[00:10:20] And now the big industry question is how to merge these two fundamental concepts to use the benefit of a CICD development concept but also have still a test and safe car.

[00:10:32] So there's a transformation ongoing in the car and turning industry where the manufacturer are changing to domain controllers and to super computers to replace these hundreds of ECUs that usually are existence in a car.

[00:10:46] So you centralize when you're going to more powerful fewer domain controllers in a car to steal the electronics.

[00:10:53] And that of course enables you to do new things faster and more efficient because you have less components to think about.

[00:11:02] And it's a fundamental transformation from a classical separation of building the car and then operating the car.

[00:11:11] So this is really difficult to separate or it's not that clearly separated anymore.

[00:11:16] It's more blurry because the capabilities to update functions in the car with auto over the air updates then enables of course that you can bring new features to the existing car that you can fix problems over the air is more easy than calling back to dealer.

[00:11:35] But especially in the dimension of a be a does actually then this CICD is a discontinuous integration and continuous development drives the need for data pipeline.

[00:11:47] So it's not a one time project where you develop a car and then you test it and then it's released and you don't see it again if you're at an a does function, then maybe you need to update that a does function because something changes.

[00:12:01] And then the big question is if you have now a life lead that is out in the field where actually you have real customer that's running the cars and operating it.

[00:12:11] And then you have to fix something I mean how do you re homologate the car on the vehicle and the fleet.

[00:12:17] How do you push new functions out to make sure actually safe and tested. And that's that is something fundamental change in the industry, where CICD becomes a component or concept that the R industry braces bring new features and up.

[00:12:30] Right features into the cars so in a CICD pipeline time is of essence. So it's quite important you have this loops fast run through so you can improve all the time.

[00:12:43] And that means actually that we need to move away from your project more in the data pipeline dimension.

[00:12:50] So it's sort of one off it's a continuous process and this means you will have continuously data coming and that's what we talk about the data pipeline.

[00:12:58] So time the processing time will be or is a differentiating factor because we have continuous loop and every loop takes you weeks.

[00:13:08] Of course, it's a very slow process so when the idea is to have iterations faster so therefore time is very important and the processing time is very important.

[00:13:17] So again, what do you do if you have a life leaf for example and then the ETA's function has an issue. What do you do if you stop it?

[00:13:28] I mean of course you want to repair it as fast as you can for CICD and then you want to update it and and on the gate and maybe in a kind of way and then go back to operations in development.

[00:13:41] You can space you can stop the development.

[00:13:44] You test invitations right in a life fleets scenario where you have operating cars running around.

[00:13:49] You cannot do that and therefore time is very, very important so classical car development programs.

[00:13:56] You have months and years to develop a function and tested in a CICD real world life fleets in our you you have hours or days to do the same amount.

[00:14:07] And of course this results decide time cost quality.

[00:14:12] There can only be achieved with a high degree of automation so actually automation is the key enabler to have a fast turnaround time that's also affordable and produces a guarantee quality.

[00:14:26] And without automation I don't think you can really achieve that level so you need to build pipelines that are up and ready to work with high level automation and then you come through the new data and you go through the automated process.

[00:14:38] And then you can iterate in a very fast way understand AI will be part of the fourth annual ICA summit set for May 22nd and 23rd in Frankfurt, Germany.

[00:14:49] The ICA is short for innovation connectivity and autonomous eight as agenda topics for the 2024 edition of the ICA summit include regulatory frameworks and safety protocols next generation sensing technologies AI machine learning and quantum computing for real time decision making among other topics.

[00:15:11] And controllers and centralized computing the role of V2X technology and cyber security and an era of connected vehicles will also be on the 2024 ICA summit agenda.

[00:15:24] If you plan to attend you will see Jurgen in the understand AI team.

[00:15:29] And if multiple activities planned so one activity is available speech about the acceleration of a data's project by leveraging CICD or data equipment development.

[00:15:40] So we'll talk a little bit more in detail on this aspect than how we can help the industry to make that happen.

[00:15:46] We also demo our products to share how we can support it in real life scenarios.

[00:15:51] And then last but not least we will have a team of experts there that can meet and discuss topics its solutions to real world problems.

[00:15:59] So we're really looking forward to a good industry exchange networking event on the challenge ends of the data programs and the discussion how to overcome that.

[00:16:08] For more information on understand AI the CICD tool chain and the 2024 ICA summit see the links in the show notes.

[00:16:17] AutoVision News Radio is available on Spotify Apple Podcast, Podbean and more.

[00:16:23] In Detroit alongside Jurgen Donnis I'm Carl Anthony AutoVision News Radio.