A football rolls into the road. A perception system labels it: ball. A more useful system wonders what may arrive next. Perhaps a child, running fast and looking nowhere near the traffic. Lei Xu likes this example because it captures the distance between seeing an object and understanding a scene. Cars have become quite good at the first task. He has built his career around the second.
Xu is a computer scientist with the résumé of someone repeatedly drawn toward the moment when an algorithm meets a physical limit. He studied computer science at the University of Science and Technology of China, completed a doctorate at the University at Buffalo, and moved through research posts before joining Qualcomm in 2012. There, computer vision had to perform in real time on embedded hardware. Memory was finite. Compute was finite. The object in front of the camera was not obliged to wait.
At Tesla, where he joined the Autopilot effort, the constraints acquired wheels. Xu led deep-learning work for Tesla Vision, building a network that detected vehicles and lane markings and replaced the Mobileye vision system used in Tesla’s first-generation products. The work reached the 2016 Model X. It was research with a factory deadline, a useful education for a founder who would soon insist that autonomous-driving technology should become a product rather than remain an impressive clip.
A startup built backward
Xu founded Nullmax in Silicon Valley in 2016 and established its Shanghai operation the following year. He took all three titles that make rest difficult: founder, chief executive and chief technology officer. The company would supply driving software rather than build a car. That choice made flexibility central. Every automaker arrives with a different vehicle, sensor set, chip, cost target and schedule. A system that only loves one immaculate hardware configuration is a fussy guest in an industry built on platforms.
Nullmax’s answer was software architecture. Its middleware separates driving applications from the hardware below them, while a common data system feeds improvements back across configurations. The company has shown products on low-compute chips for basic active safety, mid-range systems that combine highway driving and parking, and higher-compute platforms for urban assistance. Xu describes it as a technology tree: different branches run on different silicon, but learning can nourish the same trunk.
The portable intelligence loop
This architecture is an engineering decision with a social consequence. If useful driving assistance demands the costliest sensor rack and the largest computer, it remains an option on expensive cars. Xu has long treated economics as part of the technical problem. In a 2019 interview, he compared costly automation with an imaginary world of prohibitively expensive smartphones: without affordable hardware, services such as WeChat could never become ordinary. The analogy has aged better than many robotaxi forecasts.
“Data quality is much more important than quantity.”Lei Xu, on training autonomous systems
The rare frame
Autonomous-driving companies collect an ocean of road footage. Xu is preoccupied with the droplets that still teach the system something. When a model is young, ordinary scenes are useful. Later, another thousand uneventful highway miles may add little. The valuable data migrates toward the strange edge: unusual construction, an erratic neighboring car, a small object in an unexpected place. The better the system becomes, the rarer its useful lesson.
That scarcity helps explain Nullmax’s work with AI-generated visual data. Dangerous or peculiar scenes are difficult to encounter safely and on command. A generator can vary distance, weather, traffic and camera perspective while preserving a scenario the engineers need to study. Xu does not present virtual scenes as a replacement for the road. Nullmax says real-world testing remains the core of its validation, with simulation as a supplement. The point is control: ask the software the same difficult question in a hundred slightly different ways.
The football now looks less like an anecdote and more like a curriculum. Detect the ball. Infer the possible child. Anticipate motion hidden outside the frame. Slow down before certainty arrives. Xu once said an autonomous system needs not only a high IQ but something like emotional intelligence: an ability to read the situation and avoid creating danger for passengers. Engineers are not ordinarily accused of giving cars feelings. Here, the phrase is a compact description of contextual judgment.
Vision, without a religious war
Xu argues that cameras set a high ceiling for driving intelligence because they capture dense information about color, texture and movement at useful frame rates. Lidar offers direct distance measurements and can help a system reach deployment faster. He does not reject it. His more careful position is that a car should not depend on lidar to compensate for weak visual understanding. The distinction matters. Sensor debates become theatrical when framed as loyalty tests; a production engineer has to care about performance, cost, packaging and compute at once.
Cameras create their own burden. An eight-megapixel video stream is rich precisely because it is large. The algorithm must extract what matters while running on an embedded computer with fixed memory bandwidth and supported operations. Xu’s Qualcomm years show through when he discusses this. TOPS, the industry’s preferred shorthand for computing power, cannot explain the whole machine. The useful question is whether a given model can execute quickly, consistently and within the car’s resources.
The product rule
Start with the capability the vehicle needs. Work backward to software, sensors and hardware. Then make the commercial system, not merely the demonstration.
This is where Xu’s public persona becomes clearest. He has been described as pragmatic and low-key, and his language repeatedly returns to deployment: a function must create value now while contributing to a more capable system later. His 2019 formulation separated the road ahead into two legs. Driver assistance could generate revenue and data in the present. L4 autonomy, in which no driver is required within defined conditions, remained the destination. The sequence was less romantic than a leap and more survivable.
There is continuity in the people around that plan, too. Nullmax co-founder and operations chief Justin Song also came from Tesla, where he worked on Autopilot and infotainment. Desay SV became an early strategic investor and production partner. More recent relationships with chipmakers including Renesas and Black Sesame Technologies extend the same pattern: the company sits between algorithms, silicon suppliers and carmakers, translating among groups that measure success differently. A researcher may celebrate accuracy. A chip engineer watches latency and memory. An automaker wants a reliable launch at an agreed cost. Xu’s job is to make those definitions occupy the same vehicle.
“Safety is the bottom line we can never ignore.”Lei Xu at Auto Shanghai 2025
The honest boundary
In 2025, Xu offered a sentence that deserves to sit beside every driver-assistance feature list. Products should have clear safety boundaries, he said, and companies should tell users the designed scenarios without exaggeration or misdirection. This is not a decorative principle. Partial automation occupies a dangerous rhetorical zone: capable enough to amaze, limited enough to require attention. A precise boundary is part of the interface.
Nullmax’s recent work shows the two speeds of Xu’s plan. The company has pursued automatic emergency braking and production ADAS projects on varied chip platforms. It has also developed Nullmax Intelligence, an end-to-end architecture intended to combine perception, prediction and planning without relying on high-definition maps or lidar. A partnership with Renesas addresses production systems; an agreement with a European automaker points toward an L4 concept vehicle. The near road pays for the far road, but both use the same eyes.
The funding tells a similarly measured story. Nullmax announced a $10 million C1 round in December 2025 from an automotive-grade chip company. Capital, silicon and software are unusually literal companions here. The company’s ambition stretches toward passenger transport, cargo movement and robotics, areas where vision-centered embodied AI would have to read physical scenes and act within them. Yet the useful unit of progress remains smaller: one more configuration supported, one more rare case understood, one more safety limit stated plainly.
Xu has a Chinese phrase for the posture: be the minority among the majority; be a pioneer, not a martyr. It is an amusingly sober motto for an industry that once promised driverless fleets by Tuesday. The first half preserves technical conviction. The second remembers payroll, production schedules and physics. Between them sits the company he named Nullmax, traveling from nothing toward a maximum it does not pretend to have reached.
The football rolls. The camera sees it. The model considers what the frame does not show. Somewhere in that pause, before the car moves, lives Xu’s entire project: perception becoming judgment, research becoming product, and ambition learning to brake.