As Tesla’s most advanced driver assistance software (ADAS) is becoming an immediate threat due to its impending arrival in China, major Chinese electric vehicle makers and auto tech companies are hurrying to pivot their strategies towards a more pragmatic yet challenging approach to developing similar offerings. Although Tesla’s rivals have for years been looking to cut out expensive components and master the newest artificial intelligence models, the game seems to be different this time.

Both NIO and Xpeng Motors are now embracing the so-called computer vision approach, championed by Tesla, hoping their upcoming models will achieve human-like downtown driving in cities via the use of fitted cameras and radar, rather than more expensive laser sensor units. Sinpro.ai, a supplier to NIO of ultra-high-resolution four-dimensional (4D) imaging radar, said it is prepared for delivery later this year with an annual capacity of 800,000 units. Tesla has reportedly replaced radar sensors on some models after years of attempting to remove them.

“It will be a difficult task for Tesla’s Full Self-Driving (FSD) software to deal with scenarios in China where it is common that a large number of electric scooters are usually ahead in the same lane with motor vehicles,” Li Liyun, vice president of autonomous driving at Xpeng, wrote on June 27 on Chinese Twitter-like microblogging platform Weibo. Xpeng is planning to remove lidar from its upcoming sedan, scheduled for launch later this year, local media has reported.

Xpeng and NIO are also following Tesla’s suit by transitioning to an “end-to-end” autonomous driving method after using modular-based neural networks that are heavily reliant on explicit coding. Meanwhile, more traditional automakers are turning to domestic tech companies for help, such as Huawei and DJI, to catch up with the latest AI trend. Despite big challenges pressuring the industry, some early movers have the potential to compete with the US pioneer, Liu Guanghao, partner at Shanghai-based venture capital firm Befor Capital, told TechNode.

READ MORE: China opens door wider to Tesla as local giants disrupt the EV sector with AI-defined vehicles

Vision-based approach

A break from previous strategies of using expensive chips and sensors to enable ADAS capabilities, the latest approach centers on reducing the cost of components in hopes of making more room for further price cuts. Many now have their eyes on the use of radar, ditched by Tesla in 2021 due to limitations in identifying stationary objects with low image resolution, as some parts makers now said it is coming close to lidar in terms of performance – but at a lower price tag.

“There is more overlap between lidar and our sixth-generation radar systems as we significantly improve the resolution,” Juergen Brandl, Head of Market China, Business Area Autonomous Mobility at Continental Group, told TechNode. “Radar could soon see through [objects] but lidar has some problems with the distance especially in difficult situations like fog and rain.” The German firm’s newest front-facing radar boasts a detection distance of 280 and 140 meters (174 and 87 miles) for vehicles and slow-walking pedestrians, respectively.

Advanced radar solutions like this also create three-dimensional point cloud datasets like lidars, helping automakers and developers move towards fully end-to-end models with raw data collection from multiple sensors to train their self-driving systems. “We can play a big role in the development of Level 2 plus ADAS systems in China,” Brandl said, adding that the company’s product is below the price of a lidar unit with “very good“ output in terms of point cloud data.

Some disagree however, saying the technology is still at an early stage. Production of Continental’s latest-gen radar started early this year and the company began delivering the world’s first 4D radar in 2021, which detects an object’s vertical information in addition to distance, direction, and relative velocity, generating more dense point clouds than a contentional radar. Meanwhile, major Chinese lidar makers have consistently enhanced the performance of their products and lowered the prices to just over RMB 1,000 ($137.6) per unit in recent years.

The key is whether 3D/4D radar could prove to be a more “cost competitive” option compared with lidar, said Liu. “I believe both [radar and lidar] have their special advantages and disadvantages,” Brandl said. ”I think time or the market will tell whether we need both of them or one solution only.”

Either way, prospects for early players are bright. Momenta, a self-driving car company backed by General Motors and Toyota, expects the bill of materials, or total cost of components, for the City NOA (Navigation on ADAS) function to be slashed to RMB 4,000-5,000 from the current RMB 7,000-10,000 in the next two years. The intent is to survive an unprecedented price war in China that has been ongoing for more than a year.

End-to-end AI

Chinese carmakers used to brag about their coverage of cities where their assisted driving software is said to enable cars to handle on-ramp to off-ramp driving, automatic highway lane changing, and congested streets. However, many are now pivoting their focus to provide a more human-like driving experience and a full end-to-end AI model is now seen as key to winning the battle.

Described by Tesla chief executive Elon Musk as “basically photons in and controls out,” such end-to-end neural networks play an integral role in a vehicle’s decision-making process, taking raw sensor data as input and producing control actions as output. This contrasts with the conventional approaches that see each functionality, from perception to planning and action, developed individually using rule-based designs, which are often inadequate  in addressing the vast number of scenarios that occur on the road, a team of researchers said in a recent report.

The result is that people still feel their cars pilot themselves inhumanly even when kitted out with some of the most cutting-edge ADAS on the market, partly because human driving behavior tends to be consistent rather than discrete. It is very difficult for the current AV (autonomous vehicle) stack to make coherent, long-term decisions, said Wu Xinzhou, Nvidia’s vice president of automotive at its annual developer conference GTC in March.

Recent surveys have shown automakers that customers are generally dissatisfied with existing ADAS functions. Nearly half of respondents take the controls 1-2 times per 100 kilometers (62 miles) as the city NOA functions do not react appropriately, while others make more frequent interventions, according to a recent survey compiled by China’s Gaogong Industry Institute. “When the driver has to interfere pretty often, then you cannot say this is autonomous driving,” said Brandl.

Xpeng is aiming for less than one intervention per 1,000 kilometers in major traffic areas in China, CEO He Xiaopeng announced early this year, without giving a timeframe. This was followed by a new software update for its XNGP ADAS in May enhanced by the first end-to-end AI model for production vehicles in China, according to the EV maker. NIO reshuffled its autonomous driving department recently, rolling up its perception as well as planning and control teams into a single group with a focus on new AI models, Chinese media outlet LatePost reported on June 19.

It requires huge amounts of data, for example, millions of video clips, to train AI systems, as well as deep pockets and access to AI chips. Musk told investors in April that his company will increase the number of Nvidia’s flagship AI processors it uses from 35,000 to 85,000 by the end of this year. He wrote in a post on X earlier that month that the investment in training computers, gigantic data pipelines, and video storage will be well over $10 billion cumulatively this year.

Such major investment and effort is not something all companies can handle however. “It would be so hard for most traditional automakers to do this by themselves. The best way is to pick a supplier,” Liu said.

The entry of Tesla’s FSD into China may feel like a new challenge therefore, but it may also coincide with a new era of partnerships around self-driving technologies.

Jill Shen is Shanghai-based technology reporter. She covers Chinese mobility, autonomous vehicles, and electric cars. Connect with her via e-mail: jill.shen@technode.com or Twitter: @jill_shen_sh