成本与算力压力推动 ADAS 从端到端转向模块化智能体 AI

BARCELONA, Spain — Automotive executives and AI developers are reassessing the plans to bring automated driving systems to everyday vehicles. Rising semiconductor and memory costs are impacting the industry and affecting the deployment of Level 3 and Level 4 autonomous driving systems.

At AutoSens Europe 2026, Duong-Van Nguyen, global deputy CEO at VinFast, a Vietnamese EV maker, and Igal Raichelgauz, founder and CEO of Israeli AI company Autobrains, discussed the financial and technical challenges of autonomous driving systems during a session titled “From Monolithic End-to-End to Agentic AI: An OEM View on Affordable, Scalable ADAS.”

Moderated by Stav Shvartz from AutoMobility, the conversation focused on moving from complex, end-to-end neural networks to modular AI agents that can help lower production costs and reduce data needs.

Rethinking monolithic AI

Although some media coverage and industry projections suggest that fully autonomous vehicles are nearing mass-market adoption, real-world deployment of Level 4 driverless vehicles remains limited. “The reality of [autonomous] driving problem is not solved yet,” Nguyen stated during the panel. “It is partially solved in some very well-defined conditions or in areas supported by dedicated infrastructure, high-definition mapping, and continuous connectivity.”

To bring more automation to consumer cars, automakers have turned to end-to-end deep learning systems. Companies such as Tesla and BYD use these models to map sensor inputs to driving decisions and vehicle-control commands, including steering, acceleration, and braking. But handling rare situations can require much more computing power, pushing hardware costs beyond what most OEMs can afford.

“When you face different situations in the real world, the model gets bigger because you need to have more resources to compute and understand complex corner cases,” Nguyen said.

This creates tough choices for carmakers who want to keep prices reasonable and stay within the limits of their onboard electronics, he argued.

Shift to agentic architectures

To address the computing challenge, Raichelgauz introduced an approach based on agentic AI. He described three generations of driver-assistance technology. The first relied on lots of sensors and rule-based robotics, while the second used large neural networks that cut down on sensors but needed much more computing power.

Raichelgauz said the industry is now entering a third generation focused on specialization. Instead of running one huge neural network all the time, agentic AI driving breaks down tasks into smaller, specialized software agents. Each agent handles a specific driving job or situation and runs only when needed.

“Instead of solving autonomous driving as one big problem, try to break down the complexity into many similar problems where each is addressed by a specialized driver agent,” Raichelgauz said.

AutoSens 2026_From Monolithic End-to-End to Agentic AI, An OEM View on Affordable, Scalable ADAS_EE Times — From left, Stav Shvartz of AutoMobility, Duong-Van Nguyen of VinFast, and Igal Raichelgauz of Autobrains take part in the “From Monolithic End-to-End to Agentic AI: An OEM View on Affordable, Scalable ADAS” session at AutoSens Europe 2026 in Barcelona. (Source: Pablo Valerio | EE Times)

He compared the approach with the way the human brain allocates resources. The human brain uses about 20 watts of power because it only uses the resources needed for each situation, not all at once. Raichelgauz explained that applying this idea to vehicle software lets agentic systems run on much less computing power without losing performance.

Because agentic systems train each specialized model separately, automakers also need much less training data—two to three times less, according to Raichelgauz, than they would for large, unified models. This can greatly cut development costs.

Financial and testing challenges for automotive OEMs

Nguyen said that rising costs for data infrastructure and testing are major challenges for making autonomous driving sustainable. “The cost of data collection and AI development centers is unbelievable,” Nguyen noted. “It would cost you roughly $700 million just to start, and then storing the data you cannot throw away will cost $30 million to $50 million per year just to maintain it.”

Validating systems and meeting regulatory requirements create additional engineering challenges, he argued. Level 3 systems, in particular, require greater redundancy, including backup computing and sensing capabilities.

But ensuring that two redundant computing systems communicate and work together creates millions of possible test cases. Checking all these interactions across many rare situations can take years of both real-world and virtual testing.

“Most companies nowadays can manage a single-chip operation very well,” Nguyen said. “But if I put a second chip and ask how these two talk to each other, the arbitration test cases between them can go to millions of cases, taking years to test.”

Nguyen said that simulations still can’t fully replace real road testing. Virtual tests can check perception and object detection, but they don’t fully match real vehicle behavior and control.

Consumer expectations

During the Q&A, EE Times asked the panel how autonomous systems perform in areas with limited map coverage or in low-visibility conditions, particularly as drivers new to autonomous driving may expect their vehicles to handle difficult situations such as night driving or overtaking on poorly lit rural roads.

Raichelgauz replied that the main goal for consumer autonomous systems is to give drivers back 20 to 30 minutes of their daily commute, either supervised or unsupervised. Instead of making aggressive systems, computer vision-based autonomy should act with the same caution as human drivers, he argued.

“We don’t necessarily want the car to be superhuman or drive super-fast because it has better sensors,” Raichelgauz said. “When visual cues are limited, the system adapts, slows down, and acts with the same caution as a human driver, ensuring predictable behavior alongside human motorists during the extended transition period.”

Commercial deployment horizons

Nguyen said that, given current financial and regulatory pressures, carmakers are likely to focus on gradual steps toward autonomy. He expects the next two to three years to center on Level 2+ and Level 2++ systems, which provide automated driving with human supervision without the full costs and risks associated with unsupervised Level 4 systems.