VinFast recently participated in FISITA’s World Mobility Conference at the Palau de Congressos de Catalunya (June 3-5), where Dr Duong-Van Nguyen, the company’s global deputy CEO for ADAS/AD, gave a plenary keynote titled ‘Industry Disruption – The VinFast Perspective’
His presentation covered AI in scalable manufacturing design, AI sensor fusion and end-to-end AI (deep learning) in autonomous driving, as well as megatrends in EVs, connectivity and digital transformation, large language models, sustainability, green manufacturing and shared future mobility. The company also plans to launch a robotaxi for smart cities in 2027. ADAS & Autonomous Vehicle International spoke with him after his presentation to find out more.
It was interesting to hear about end-to-end AI in your presentation. Is this something that VinFast is already working on? Are other OEMs working on end-to-end AI?
Yes, we indeed have already been working on solutions using end-to-end AI, as it shows significant potential to enable more fully automated applications, ranging from manufacturing to autonomous driving. Nevertheless, it must be clear that AI at the modular level has been used extensively in many domains within automotive mass production, whereas end-to-end AI is still in a very early stage, especially for autonomous driving.
Today, AI still relies on brute force, with the need for a tremendous amount of data. Thus, end-to-end AI requires even more data for training. This leads to the long-tail challenge in enabling autonomous driving, which is not favorable at all for mass production. In other words, it resembles continuous development and demonstration rather than a true mass production approach. You can observe that Tesla or a Chinese OEM has to continuously collect data and release frequent software updates to address rising on-road issues.
In the past, mass production meant producing a set number of vehicles with minimal maintenance expectations — that’s how profitability was achieved. With end-to-end AI, you must continuously monitor infinite on-road events, which results in excessively high costs at a global deployment scale. Maybe Tesla’s AI data center is used for multiple purposes like automotive, robotics, etc., and they can find investors to fund that. But imagine a company that solely manufactures cars and wants to implement end-to-end AI – the development and maintenance costs would be too high to sustain profitability. I suppose that might be one reason it is difficult for Chinese OEMs to expand globally with their advanced autonomous systems.
So, end-to-end AI is not economically viable on a company level?
For us, we can do it. We have automotive, real estate, robotics, universities, hospitals, hotels, a taxi company — all of which can share the cost of an AI center. We will make it anyway.
When discussing end-to-end AI in your presentation, are you talking solely about end-to-end AI within the vehicle itself – within all the sensors and systems or sensor fusion – or about the development process as well?
Actually, it is about everything. In fact, it may be easier to apply it to process, documentation and manufacturing than to the vehicle itself. Despite the high cost, technically, end-to-end AI is relatively straightforward to deploy, even within the vehicle. We just need to prepare a closed-loop toolchain for data collection, annotation, training and execution, all of which are very much supported by current AI technologies and infrastructure.
For ADAS and autonomous driving, without end-to-end AI, we often use AI at the component level, such as perception, fusion and path planning. This allows us to separately control the quality of each component within the autonomous system and incorporate rule-based models to ensure safety and reliability. End-to-end AI, on the other hand, works by training a single large network with low-level fusion inputs (for example, data from cameras, lidar, radar, etc.) and outputs at the path planning or even vehicle control level (torque, steering wheel, etc.). That’s it. It sounds very simple, right?
You said in your presentation that you need more data – is this an issue?
We need a lot more data. The amount of data required for end-to-end AI increases exponentially. With distributed AI, we apply AI to collected data from specific sensor sets and generic vehicle information (like CAN, GPS, etc.) that are not highly coupled with a particular vehicle. This allows us to reuse data and models with minimal adaptation, even across new vehicle platforms. Simulation also plays a major role in such training.
However, for end-to-end AI, simulation is far less useful. Synthetic data cannot perfectly reflect real-world data, so there’s no guarantee that a simulated solution will perform correctly in reality. As a result, a vast amount of real data is required to validate performance.
End-to-end AI also encodes vehicle dynamics coupled with driver behavior, making it difficult to transfer models to other vehicle types without significant effort — perhaps 40-50% of the original development effort. This level of cost is too high for scalable mass production.
How can you get more data? Would that require more data sharing around the world?
It seems that Chinese OEMs have the best access to data sources, enabling their advanced autonomous functions through the end-to-end AI approach. Tesla has also managed to collect a significant amount of data in recent years, giving it the capability to pursue end-to-end AI. Still, the cost remains very high — it’s a game of hundreds of millions, if not billions.
I would say that everyone needs to follow the path of end-to-end AI, as it provides the easiest way to showcase impressive products. Over time, data will accumulate. In the past, we never imagined that someone could display live traffic conditions around the world, yet now we have that through apps like Google Maps. The same will happen with AI — it’s just a matter of time. Of course, forcing people to share data is nearly impossible. Rumor has it that data sharing exists in China, but somehow the process will evolve eventually.
You said that AI doesn’t care about your stack – could you elaborate on this?
The way AI solves problems differs from human logic. People claim AI is intelligent, but today’s AI is not – at least not in the human sense. The AI models we currently use, like CNNs, R-CNNs, LSTMs and transformers, are essentially memory-driven models that calculate the likelihood of outcomes based on existing data. For instance, given a word, the model predicts the next most probable word or scenario. That’s why it’s called a language model – because language itself functions probabilistically. We design networks to allow feedback, but we don’t fully understand how that feedback is formed.
Many exaggerated claims about AI are made just to generate hype. Years ago, people said AI would learn your shape and behavior to detect and analyze human activity. But in truth, we don’t know what it’s learning. Even now, many boast about AI’s great performance and failures are dismissed as “corner cases”. Yet do we truly understand what those are? No. Sometimes, it’s simply a misdetection of a person wearing an unfamiliar color not included in the training set. So, AI corner cases are not “rare cases” in human logic, they are just statistically unlikely events based on the dataset.
Still, end-to-end AI is impressive in many ways. I never imagined we could one day have an automotive factory running almost autonomously.
You say that natural language processing is just a natural evolution of things, or what you’re doing anyway, when everyone’s passing around all these trendy buzzwords …
I come from a generation that experienced the rise and fall of AI hype in the 1990s, followed by another wave from 2012-2018, and now from 2021 onward. The fun part is that our generation spent a lot of time crafting sophisticated human rule-based models to tackle complex problems. Then one day, we saw a trivial network outperform them. It kept improving, and we pivoted toward that field to make our contributions. So, for our generation, it makes perfect sense. We are excited to see the arrival of a new generation of GenAI that can truly perform reasoning and improve its generalization. The younger generation now benefits from AI tools without needing to understand the underlying math. They often perceive it as something complex and magical. But in reality, techniques like gradient descent and backpropagation are among the simplest optimization algorithms around.