Beyond Algorithms: Where Does Autonomous 슬롯사이트 온카판 Competitiveness Diverge?

There is a growing consensus that 슬롯사이트 온카판 has become more important than algorithms in the autonomous driving industry. From your perspective in the field, what is the biggest challenge in current development?
Dahlstrom: I agree that 슬롯사이트 온카판 is positioning itself as the key differentiator, especially in machine learning approaches rather than entirely new technologies. As Andrew Ng (Stanford Professor and founder of DeepLearning.AI) has noted, most algorithms today are essentially accessible to everyone. You can use algorithms nearly identical to those used by Google or NASA. The real challenge lies in whether you can train those algorithms into production-grade models. Ultimately, that depends on the availability of 슬롯사이트 온카판.
Autonomous driving does not have the kind of 슬롯사이트 온카판 environment available to large language models. LLMs can rely on the internet as an immense 슬롯사이트 온카판 source, but autonomous driving has no equivalent at that scale. Collecting large amounts of 슬롯사이트 온카판 easily and at low cost is still difficult. That is why building a 슬롯사이트 온카판 pipeline alone is not enough. The real challenge is creating a continuous 슬롯사이트 온카판 flywheel. What matters is not simply having more 슬롯사이트 온카판 than everyone else, but securing the right 슬롯사이트 온카판 at the right scale. In the end, the advantage will go to whoever can iterate faster.

 


The way OEMs and AI companies build their 슬롯사이트 온카판 infrastructure seems to have changed rapidly over the past few years. How have customer requirements evolved?
Dahlstrom: We are seeing a very interesting shift right now. Many companies are moving away from traditional AD architectures. The conventional structure usually follows a hierarchical path: perception - prediction - planning. However, there is a trend toward larger, integrated models, such as End-to-End AI or Vision Language Models (VLM).
Both approaches have their pros and cons. But with VLM, new data requirements emerge. Previously, we taught models how to identify objects, and then used that information to understand the scene and decide on appropriate actions. In VLM, however, "reasoning" is embedded within the model. Put simply, the focus is shifting from "What" to "Why."
This shift also affects 슬롯사이트 온카판 annotation. Tasks that machines already do well - such as describing object shapes or geometric properties - are left to the machines. Humans, meanwhile, take on the role of providing feedback on what is important in a given situation and which factors influence decision-making.


As the discussion moves toward behavioral understanding and Physical AI, how will 슬롯사이트 온카판 requirements change?
Dahlstrom: This is a very hot topic of debate right now. VLM is being used not only for 슬롯사이트 온카판 curation but also as an actual deployment model, sometimes referred to as a Vision-Language-Action (VLA) model.
The goal of this approach is to leverage the vast world knowledge inherent in language models to solve the "long-tail problem." While much changes regarding data, some things remain the same. These models still require human feedback to align model behavior with the expectations of human drivers or passengers.
However, the nature of human feedback is changing. Previously, humans labeled "what" was around the vehicle - using bounding boxes, for example. Now, we need explanations for "why." This means feedback that explains which elements are important in a specific situation and the logic the vehicle should follow. A study that has recently made a significant impact in this field is NVIDIA  s Alpamayo paper (a study on VLM-based driving scenario reasoning and data alignment).





 
Many companies try to internalize their 슬롯사이트 온카판 pipelines. Why is a specialized platform like Kognic still necessary?
Dahlstrom: 슬롯사이트 온카판 requirements, like all other requirements, are never static - especially in a cutting-edge field like autonomous driving. No company can sit down on day one and create a plan that remains perfectly fit for three years, or even six months. If you try to build every element of the 슬롯사이트 온카판 pipeline internally, you must manage not only simple maintenance but also unpredictable demands for change. Total control is an advantage, but it comes at a cost, and that cost is difficult to predict.
Traditionally, automotive companies dislike this uncertainty. OEMs and Tier 1s usually prefer fixed cost structures over time. Unless you have nearly unlimited funds, internalizing everything is a very difficult decision.
The trend we see now is a compromise strategy. Many automotive companies are adopting modular and flexible strategies. They maintain control and analytical capabilities over the entire 슬롯사이트 온카판 pipeline internally but bring in individual components - such as the annotation engine - from external providers. They also maintain options to switch providers by using multi-cloud, API-based structures, and flexible contracts if future requirements change. Kognic supports multiple companies simultaneously, allowing us to spread costs and create economies of scale. As a result, we can provide competitive pricing with relatively low risk.


Autonomous driving 슬롯사이트 온카판 annotation is not just simple labeling; it is closely linked to safety. How does Kognic define and manage 슬롯사이트 온카판 quality?
Dahlstrom: Safety is a paramount theme in this field. While development speed and market pressure are increasing, I hope safety remains a core essential requirement. Most safety engineering happens at a higher abstraction level than what we directly intervene in - for example, at the functional or system level. However, we are naturally responsible for 슬롯사이트 온카판 quality.
To this end, the Kognic platform and processes include multiple layers of Quality Assurance (QA). These include automated "sanity checks" to prevent labeling errors from entering the workflow, project management analytics to identify the root causes of error types qualitatively and quantitatively, and Bayesian probability-based KPI quality statistics.


What is the most difficult part of autonomous driving 슬롯사이트 온카판 work? Specifically, what problems arise in multi-sensor fusion environments involving LiDAR, Radar, and cameras?
Dahlstrom: Test vehicles or production vehicles are usually equipped with various multimodal sensors: rotating LiDAR, flash or low-scanning LiDAR, cameras, and Radar. These sensors are mounted in different positions, have different scanning methods, and are often not perfectly synchronized in time. Consequently, no two sensors observe the same object at the exact same location and timestamp.
We have put a lot of effort into solving this. The Kognic platform supports sequence-based multi-sensor 슬롯사이트 온카판 processing. It is designed to maintain speed and cost-efficiency while compensating for sensor modality differences, ego-motion, and time lags.


 


AI-based auto-labeling technology is advancing rapidly. How do you see the roles of humans and AI evolving in the 슬롯사이트 온카판 generation process?
Dahlstrom: There has been a major shift here. When we first started the company, labeling was almost entirely manual. We focused on creating tools to help humans work faster. Now, however, most customers have very powerful auto-labeling algorithms. These algorithms run offline in high-performance computing environments and can analyze recorded log 슬롯사이트 온카판 chronologically backward and forward, often outperforming the models installed in the vehicles.
However, they are still not at a level of total trust. The human role is still necessary, but its nature has changed. In many current projects, the human role is essentially "QA-centric." We process model predictions through additional automation on our platform and then direct human attention to objects or frames with a high probability of needing correction or approval. We call this "Model & Human in the Loop." Ideally, the goal is for humans and models to collaborate efficiently, with humans investing time only in the areas where the model is not yet proficient.


 
Simulation or synthetic 슬롯사이트 온카판 is used to supplement the lack of real-world driving 슬롯사이트 온카판. What are the realities and limitations of this approach?
Dahlstrom: I believe synthetic or augmented data is an extremely useful resource, especially for model training. As mentioned, autonomous driving faces the "long-tail problem." There are rare but critical situations that are realistically or ethically difficult to record intentionally on real roads. However, sensor realism remains a major challenge. Camera-based photorealistic simulation is now very advanced, but generating realistic LiDAR scans or Radar RCS (Radar Cross Section) without much real-world data is not easy.
Therefore, there is much debate about using non-real 슬롯사이트 온카판, especially in the validation stage. Proving that a complex synthetic scenario is sufficiently similar to the real world can be as difficult as the problem we were trying to avoid in the first place.


Then how can long-tail scenarios be managed?
Dahlstrom: This is a practical challenge. However, looking into and understanding the data we already have is now technically possible. Usually, we approach this through overlapping steps rather than a single method. First, we use metadata such as time, region, and CAN bus signals to narrow down the data to the desired conditions - for example, "highway driving on a rainy night." Then, we perform a quick visual check to see if the filtered results match what we are looking for.
Going further, we can use models like VLM to search for more abstract concepts - not just weather or time, but contextual scenes like "a person standing by the side of the road." Finally, a human enters to validate. Humans quickly confirm model-identified results (yes/no), increasing accuracy and using those results to further improve the model. Through this process, we can accurately determine how many long-tail cases, such as "a person on a highway on a rainy night," actually exist within our real data.


Can an environment like China  s, which secures large-scale vehicle data quickly, create a structural advantage in the autonomous driving race?
Dahlstrom: Data scale is certainly important. A certain level of scale is required for training and validation. Furthermore, if you can collect data from a vehicle fleet, the increased mileage and diversity of situations heighten the probability of capturing long-tail cases. However, data itself is not the asset. The true asset is the ability to connect that data to actual model performance improvements. To do this, you must build the "data flywheel" I mentioned earlier. Regarding China, from a limited perspective, many companies in that market seem to understand this point very well.





 
In the SDV era, OTA updates and continuous 슬롯사이트 온카판 collection become possible. How will the feedback loop between 슬롯사이트 온카판 collection, model training, and deployment change?
Dahlstrom: This connects directly back to what we discussed earlier. You can call this continuous feedback, the "Big Loop," or the "Data Flywheel." The product paradigm in the SDV era is completely different from the past. In the past, a car was in its most perfect state at the moment of purchase. In the SDV era, it is the opposite. A car must continue to improve while it is being used. This means data collection, model training, and model deployment must be connected into a single, integrated, and iterative system. This change is by no means easy for traditional automotive companies, as they were optimized for a completely different product development philosophy.


There is a perception that Kognic is related to the Volvo ecosystem. What is the actual relationship?
Dahlstrom: Kognic  s first and oldest customer is Zenseact, a subsidiary of Volvo Cars. Zenseact develops autonomous driving systems and is located in Gothenburg, Sweden - the same city as us. In fact, our offices are only about 100 meters apart on the same street. However, we have been a separate, independent company from the start. Currently, Kognic primarily serves automotive OEMs and Tier-1 companies (showing logos of Qualcomm, Zenseact, Continental, Bosch, Kodiak, ZF, Embotech, Einride, Gatik, JLR, etc.). Which company develops the actual AD stack or perception system depends on the region. In Europe, OEMs often rely on suppliers, while in Japan or the US, some OEMs prefer in-house development. We also collaborate with Level 4 companies, such as those in autonomous trucking.


 

AEM(오토모티브일렉트로닉스매거진)

---