The perception stack of autonomous vehicles must be able to perfectly detect the vehicle's environment. For this purpose, its function is compared with ground truth data. This data is automatically generated by means of annotation based on artificial intelligence and machine learning. The aim of understand.ai, a company in the dSPACE group, is to optimize the quality, implementation speed, and costs of annotations.

Jürgen, you started as CEO of understand.ai in 2021. What did you do before that? What are your impressions of the company and of the dSPACE Group?

My professional background is in the automobile industry, in the fields of EE and IT, with a focus on CASE (Connected, Automated, Shared, Electrified). I worked in technical and sales management positions, in both start-ups and large companies. The realization of new and innovative technologies in the automobile industry motivates me greatly. understand.ai is a driver of innovation and has exceptional competence and products in the fields of artificial intelligence and machine learning that we apply in the autonomous transportation environment. We have left the start-up phase and are now in a scale-up phase with a highly-motivated and competent team. With dSPACE in the background, we have a strong partner to further expand our growth. We work independently, but use synergies where it makes sense. It should also be mentioned that the understand.ai products fit very well into the dSPACE data-driven development portfolio, e.g., in conjunction with the generation of simulation scenarios.

When will we experience autonomous driving and which challenges does it present for manufacturers of autonomous vehicles?

First we need to distinguish between whether we are talking about passenger cars or commercial B2B applications. In the field of passenger cars, we see today mainly functions in the ADAS/L2 area and initial applications in the AD/L3 environment. The questions to be answered here, in addition to technology, are the costs, the approval conditions, and the question of liability which is increasingly being transferred from the driver to the manufacturer. Therefore, in my opinion, autonomous driving at Level 4 or Level 5 for the passenger car segment will still take some time. The situation in the commercial sector is different, because the operational design domain (ODD) can be more clearly defined and therefore approval and liability risks can be limited. Furthermore, the business case pays off commercially, because the costs of AD technologies can be offset by the elimination of the driver. We are already seeing the first L4 use cases in practice today.

What are the most important success factors for the efficient development of systems for automated driving?

A common prerequisite for successful implementation of autonomous driving is the quality of the data, algorithms, processes, and technologies, as well as the speed and degree of automation in the development process. Reliable tool chains that facilitate high flexibility, quality, and automation must be established. The volumes of data to be processed are huge, but the costs need to be kept under control. The development and response times also need to be shortened, in order to bring new sensors and functions into series production in a timely manner and to ensure the safety of the fleet in the field at all times.   

Automation: Decoupling the linearity between manual annotation costs and data volume.

How is understand.ai supporting its customers? What is the value for the customer?

understand.ai creates annotations for the generation of ground truth data for autonomous transport systems. Large quantities of data are required in approval-relevant application areas in order to validate the perception stack. During this process, the functionality of the perception stack is compared with the so-called ground truth. We use artificial intelligence and machine learning to fully automate the annotation creation process. The goal is to avoid manual labeling work when the validation data is being processed. The traditional annotation approach involves a very large amount of manual work to adjust and check the labels. Large validation projects can easily mean that hundreds or even thousands of employees are needed, which is very expensive, difficult to coordinate, and slow. Our automation approach enables an affordable price for large-scale validation projects, a methodically validated data quality, and also a significantly shorter project duration.

What does a typical understand.ai project look like?

We receive the data from our customers, or we establish a data pipeline, to continuously process data. If required, we then import and calibrate the data and anonymize it. This is necessary, for example, if humans need to access the data for training purposes, in order to select data. Next, our labeling robots are developed, customized, and trained. By means of an iterative process, we improve the quality of the automation to the desired level of quality. Then the data is processed and labels are automatically generated. In the final step, the results are sent back to the customer. Our platform runs as a SaaS solution in the cloud, which enables us to scale up and down as required.  

Which strategic directions can customers expect from understand.ai?

We are focusing on automation in the field of large-scale validation. We place particular emphasis on the subjects of quality, scaling, and flexibility. For us, quality means that we achieve the required data quality level, but also that we possess the methodology to demonstrate compliance of the quality. In the area of scaling, we are continually increasing our ability to adapt our systems to the increasing data volumes and to parallelize it massively. We are also automating our processes continually so that we are able to handle a large number of projects simultaneously. Our flexibility means our ability to quickly adapt to customer requirements - both from a functional and process perspective. Our aim is to be part of the development tool chain for OEMs, Tier1, B2B transport providers, and tech companies in the field of AD/ADAS.

Left: Raw data. Right: Annotated data (road markings).

What successes can understand.ai already report?

We now have a very clear strategy and an implementation plan within understand.ai and also in the context of dSPACE. We are growing very strongly and gaining more and more customer projects in our target segments. The feedback from the industry, from customers, partners, and analysts is positive across the board, and it is clear that we are developing a solution for an industry problem that has not yet been solved.

Thank you for talking with us.

ニュースレターを購読します

メールマガジンの購読希望・変更/配信停止手続き