Vehicles

From PRIMO.ai
Jump to navigation Jump to search

Youtube search... ...Google search


Artificial intelligence now touches nearly every layer of transportation, from the computer vision that lets a car recognize a cyclist to the models that forecast congestion on a city's roads. This page focuses on autonomous vehicles (AVs), which use sensors, onboard computers, and machine learning to perceive the road and drive with reduced or no human input, but it also covers advanced driver assistance systems (ADAS), traffic prediction, vehicle price prediction, camera-based surround vision, and the AI-driven factories that build modern vehicles. The field grew out of the DARPA Grand Challenges of 2004, 2005, and 2007, when university teams showed that robotic vehicles could complete desert and simulated urban courses on their own; Sebastian Thrun's Stanford team won the 2005 race with its car, Stanley. Google launched its self-driving project in 2009 with many of those researchers, and that project became Waymo in 2016.

What separates the AI approach from conventional vehicle automation such as cruise control or anti-lock brakes is that the hardest parts of driving (recognizing objects, predicting what other road users will do, and choosing a safe path) are learned from data rather than written as explicit rules. Early systems combined hand-engineered software with learned perception modules, while the newest systems train large neural networks on millions of hours of driving and borrow techniques from large language models and generative AI. By late 2026, fully driverless robotaxis were carrying paying passengers in more than a dozen U.S. metropolitan areas and many Chinese cities, driverless trucks were hauling freight between Texas and Arizona, and conditional "eyes-off" automation had been approved in Germany, parts of the United States, and pilot zones in China. Most cars on the road, however, still rely on Level 2 assistance that requires a human to supervise at all times.

Self-driving cars will enable car-sharing even in spread-out suburbs. A car will come to you just when you need it. And when you are done with it, the car will just drive away, so you won’t even have to look for parking. – Sebastian Thrun, CEO of Udacity and a Google engineer who has been called the father of self-driving cars

The stakes are large. Road crashes kill about 1.19 million people a year worldwide according to the World Health Organization, and most crashes involve human error such as distraction, impairment, or speeding. Automation also promises mobility for people who cannot drive, round-the-clock freight movement, less land devoted to parking in smart cities, and smoother traffic that saves energy. Those benefits come with real risks and open questions: how to prove a system is safe before it is deployed widely, who is liable when it fails, how to protect vehicles against cyberattacks, how to handle the privacy of camera data, and how to treat the professional drivers whose jobs may change, all of which raise ethical questions as well as technical ones.

Levels of Driving Automation

The most widely used vocabulary for automation comes from SAE International's standard J3016, developed with ISO, which defines six levels from Level 0 (no driving automation) to Level 5 (full driving automation). The levels describe who performs the dynamic driving task, meaning the real-time steering, braking, and monitoring of the road, and who serves as the fallback when something goes wrong; they say nothing about how safe or polished a system is. Each automated feature is designed for an operational design domain (ODD), the specific conditions such as road types, speeds, weather, and geography in which it is meant to work. SAE published a revised edition of J3016 in September 2026 that rewords the short level descriptors to make the differences between levels easier to understand, while stressing that those descriptors are not the formal definitions.

Level Name Who drives when engaged Who is the fallback Examples (2026)
0 No driving automation Human Human Warnings and momentary interventions such as automatic emergency braking
1 Driver assistance Human, with steering or speed support Human Adaptive cruise control; lane-keeping assist
2 Partial driving automation Human, with steering and speed support Human, who must supervise continuously Tesla FSD (Supervised); GM Super Cruise; Ford BlueCruise; BYD God's Eye
3 Conditional driving automation System, within its ODD Human, who must take over when requested Mercedes-Benz Drive Pilot; Changan and BAIC Arcfox pilot models in China
4 High driving automation System, within its ODD System Waymo, Baidu Apollo Go, Pony.ai, WeRide, and Zoox robotaxis; Aurora driverless trucks
5 Full driving automation System, everywhere System None deployed

The most important dividing line falls between Level 2 and Level 3. Levels 0 through 2 are driver support features, so the human is still driving even when a "hands-free" system such as Super Cruise is steering; Levels 3 through 5 are automated driving systems (ADS), so the system is driving whenever it is engaged. Marketing terms such as "Level 2+" or "Level 2++" do not appear in J3016 and usually describe Level 2 systems with more capable features, which is why safety researchers worry about drivers over-trusting them. Because the system is responsible for driving at Level 3 and above, liability tends to shift toward the manufacturer; Chinese guidance, for example, states that manufacturers bear primary responsibility for incidents involving Level 3 and higher systems.

How Autonomous Vehicles Work

An automated driving system has to answer four questions many times per second: what is around me, where exactly am I, what will everyone else do next, and what should I do now. AVs are a demanding form of mobile robotics and embodied AI, because the machine must turn noisy sensor data into physical actions in a shared, unpredictable environment where errors can injure people. Engineers have traditionally answered those questions with a modular pipeline of separate components, but since about 2023 the industry has moved toward larger learned models that merge several stages. Both approaches depend on the same foundations: rich sensing, large and well-curated datasets, extensive simulation, and powerful onboard computers.

Sensing and Sensor Fusion

Most AVs combine several complementary sensors. Cameras provide color and texture, which makes them essential for reading traffic lights, signs, and lane markings, but they struggle with glare and darkness and must infer depth indirectly. Radar measures distance and relative speed directly using radio waves and keeps working in rain and fog, although at low resolution, and its returns are interpreted with classic signal processing as well as neural networks. Lidar fires laser pulses to build precise three-dimensional point clouds of the surroundings, while ultrasonic sensors handle close-range parking, and satellite positioning (GPS), inertial measurement units, and wheel odometry track the vehicle's own motion.

Sensor fusion combines these streams into one consistent picture of the world, either by merging raw data early, merging detected objects late, or, increasingly, projecting everything into a shared learned representation such as a bird's-eye-view grid. The industry remains divided over how much sensing is enough. Tesla and XPeng's robotaxi program rely on cameras alone, arguing that people drive with vision and that cheaper hardware scales faster, while Waymo, Zoox, and most Chinese robotaxi operators use lidar and radar for redundancy, Mercedes-Benz says lidar is essential for its Level 3 system, and GM plans to add lidar for its 2028 eyes-off system. Hardware costs are falling quickly on both sides: Pony.ai says the materials cost of its seventh-generation autonomous driving kit fell by nearly 70 percent, and WeRide reports a 50 percent cut in kit cost.

The Modular Driving Stack

In the classic modular design, perception comes first. Convolutional neural networks and, more recently, Transformer models perform object detection, segment drivable space and lanes, classify traffic signals, and estimate the speed and heading of every vehicle, pedestrian, and cyclist, including objects that are partly hidden behind others, a problem known as occlusion. Localization then matches what the sensors see against detailed prior maps to place the vehicle within a few centimeters, which allows the software to know exactly which lane it occupies. Some newer systems reduce their dependence on high-definition maps so they can expand to new cities faster, at the cost of having to infer more of the road layout in real time.

Prediction and planning come next. Prediction models forecast several possible future trajectories for each nearby road user, often using recurrent networks, attention-based models, or graph neural networks that capture how agents influence one another. The planner then chooses a trajectory that balances safety, comfort, legality, and progress under uncertainty, a problem often framed as a Markov decision process with partial observability, and a controller converts that trajectory into steering, throttle, and brake commands. The modular approach is easier to inspect and test because each component has clear inputs and outputs, but errors can compound across stages and hand-written rules struggle with the endless variety of real-world situations.

End-to-End Models, Vision-Language-Action Models, and World Models

End-to-end driving replaces much of the hand-built pipeline with one large network that maps sensor input directly to a planned trajectory. These models are trained mainly through imitation learning on large amounts of human driving data and refined with reinforcement learning in simulation, and they often keep intermediate outputs such as detected objects so engineers can inspect what the model "sees." Tesla moved its FSD software to an end-to-end neural network in 2024, the UK company Wayve has pursued end-to-end learning since its founding in 2017, and the research model UniAD showed in 2023 that a single network could jointly handle perception, prediction, and planning. The main advantage is that performance improves with more data and compute rather than more rules; the main drawback is that a single large network is harder to interpret and verify.

Vision-language-action (VLA) models extend this idea by attaching a driving "action" output to a vision-language model, so the system can reason about an unusual scene in words before deciding what to do, a technique related to chain-of-thought prompting. NVIDIA released Alpamayo-R1, an open reasoning VLA for driving research, at NeurIPS in December 2025, followed at CES 2026 by Alpamayo 1, a 10-billion-parameter model built on its Cosmos-Reason backbone, together with the AlpaSim simulator and an open dataset of more than 1,700 hours of driving. Its successor, Alpamayo 2 Super, pairs a 32-billion-parameter vision-language model with a 2.3-billion-parameter diffusion-based action decoder and outputs a trajectory, a written explanation, and a high-level maneuver such as "yield" in one pass; NVIDIA released it for commercial use on Hugging Face in August 2026. In China, XPeng began rolling out its VLA 2.0 system to customers in March 2026, describing a 72-billion-parameter model trained on about 100 million driving clips that maps vision directly to action, and Volkswagen became the first outside customer for both the model and XPeng's Turing chip.

World models are generative models that predict how a driving scene will evolve given the vehicle's actions, which lets developers create rare or dangerous scenarios and test driving policies in closed loop without real-world risk. Wayve's GAIA series illustrates the trend: GAIA-1 scaled to 9 billion parameters, and GAIA-3, released in December 2025, has 15 billion parameters (twice GAIA-2) and a video tokenizer twice as large. Wayve reports that GAIA-3 test results track real-world driving closely and cut the rate of rejected synthetic tests fivefold, and it is working with the University of Warwick on a UK government-funded project to validate world models for safety evaluation. These tools build on foundation models and simulated environment learning, but they can still produce physically implausible scenes, so their results must be checked against real-world data before they count as safety evidence.

All of this runs on specialized automotive computers rated in trillions of operations per second (TOPS). NVIDIA's DRIVE AGX Thor powers platforms such as WeRide's GXR robotaxi and Wayve's third-generation autonomy computer, XPeng's GX robotaxi carries four in-house Turing chips delivering up to 3,000 TOPS, and BYD's DiPilot computers range from 100 to 600 TOPS. Because a car cannot carry a data center, developers shrink models with quantization and distillation, write custom compilers, and design dedicated processing units so that billion-parameter networks can respond within milliseconds while drawing modest power.

Safety, Evaluation, and Validation

Proving that an AV is safer than a human driver is difficult because serious crashes are rare per mile driven, so a meaningful comparison requires hundreds of millions of miles of data matched to the same roads and conditions. Waymo publishes the most detailed public safety data through its Safety Impact hub, which compares crashes reported to NHTSA's Standing General Order with human benchmarks for the same areas. Its June 2026 update covered more than 220 million rider-only miles through March 2026 and found 94 percent fewer crashes causing serious or fatal injuries, 82 percent fewer airbag-deployment crashes, and 82 percent fewer injury crashes than human drivers; a September 2026 update reported 95 percent fewer serious-injury-or-worse crashes and 93 percent fewer pedestrian injury crashes. A related peer-reviewed analysis in the journal Traffic Injury Prevention found a 96 percent reduction in injury-involving intersection crashes, and an insurance study by Swiss Re reported similar drops in liability claims.

These results are encouraging but carry important caveats. The analyses are produced by the company that operates the vehicles, human benchmarks must be adjusted for underreported crashes, and robotaxis operate within limited service areas that may differ from average driving conditions; fatality rates remain too rare to measure precisely. Other operators publish far less: Tesla, for example, has made safety statements about its robotaxi service but has not released robotaxi crash statistics comparable to Waymo's. Independent experts such as Carnegie Mellon University's Philip Koopman argue that safety requires more than mileage statistics, including a documented safety case, standards such as ISO 26262 (functional safety), ISO 21448 (safety of the intended functionality), and UL 4600, and transparent reporting of incidents.

Developers rely on several layers of verification and validation. Scenario-based testing on closed tracks is combined with billions of simulated miles, reinforcement learning environments in which each driving decision changes the simulation, and world-model-based evaluation, while model evaluation on public benchmarks such as nuScenes and the Waymo Open Dataset tracks research progress. The hardest problem is the "long tail" of rare events, from a mattress falling off a truck to a police officer directing traffic by hand, which requires out-of-distribution generalization. Reasoning models that explain their decisions in natural language are one attempt to make behavior auditable, connecting AV safety to broader work on explainable AI.

Security is part of safety. Researchers have shown that adversarial attacks, such as specially designed stickers on signs or spoofed lidar returns, can fool perception systems, which has motivated research into defenses against adversarial attacks and sensor redundancy. Connected vehicles that receive over-the-air updates and rely on remote assistance centers also present a larger cybersecurity attack surface, which is why UN regulations R155 and R156 now require cybersecurity and software-update management systems for new vehicles in many markets.

Regulation and Governance

In the United States, NHTSA sets vehicle design standards through the Federal Motor Vehicle Safety Standards (FMVSS), while states regulate drivers, licensing, insurance, and how AVs may operate, and Congress has not yet passed comprehensive federal AV legislation. Many FMVSS assume a human driver and require items such as steering wheels, mirrors, and windshield wipers, so vehicles built without manual controls need exemptions, which are capped at 2,500 vehicles per manufacturer per year. NHTSA introduced an AV Framework in April 2025, proposed rules in March and April 2026 to modernize standards such as FMVSS 102, 103, 104, and 110 for driverless designs, and listed eight rulemakings covering ten standards in its July 2026 agenda. On July 30, 2026, it granted Amazon's Zoox the first commercial exemption for a purpose-built robotaxi without a steering wheel or pedals, valid for two years and up to 2,500 vehicles per year, and the next day it issued an interim final rule allowing vehicles built before an exemption to qualify for commercial deployment. Separately, a Commerce Department rule issued in January 2025 restricts Chinese- and Russian-linked software and hardware in connected vehicles sold in the U.S., which shapes which AV technology can enter the American market.

China regulates AVs through national ministries and city-level permits. On December 15, 2025, the Ministry of Industry and Information Technology (MIIT) conditionally approved the country's first two Level 3 passenger models, from Changan and BAIC's Arcfox brand, for pilot use on designated roads in Chongqing and Beijing, and in October 2026 an MIIT official said a second batch of approvals was underway. China has also set a unified safety baseline for Level 3 and Level 4 systems and proposed turquoise exterior lamps to signal when an automated driving system is in control. Large cities such as Beijing, Shanghai, Guangzhou, Shenzhen, and Wuhan issue their own robotaxi permits, and Shenzhen granted its first citywide driverless commercial permit to Pony.ai and a local taxi operator in October 2025, although permit issuance has occasionally paused in some cities.

Europe and Japan have taken a more incremental path. Germany allows Level 3 driving at up to 130 km/h under UN Regulation 157 and permits Level 4 operation in approved areas under a 2021 law, and Mercedes-Benz received approval in December 2024 to operate Drive Pilot at up to 95 km/h across the 13,191-kilometer Autobahn network. The UK passed the Automated Vehicles Act in 2024 and later brought forward commercial self-driving pilots to 2026, which Wayve and Uber plan to use for a London robotaxi service, while Japan has permitted Level 4 operation since 2023 and will host a Nissan, Uber, and Wayve robotaxi pilot in Tokyo by late 2026. Across all regions, AV policy is becoming a test case for AI governance, forcing regulators, insurers, and courts to decide how to assign responsibility when software drives; see also risk, safety, compliance, and regulation, law, and government services.

Deployment Landscape

Robotaxis

Waymo is the clear leader in U.S. robotaxi deployment. It reached 500,000 paid rides per week across ten U.S. cities in March 2026, a tenfold increase from roughly 50,000 weekly rides in May 2024, with a fleet of about 3,000 to 3,500 vehicles, and in July 2026 it began fully driverless operations in Las Vegas and other new markets. The company's co-CEO has said it aims to exceed one million paid rides per week by the end of 2026 and to operate in more than 20 cities, including London and Tokyo, sometimes moving from mapping a city to paid service within months. Waymo works with Uber in some markets, and in April 2025 it agreed to explore a partnership with Toyota that could bring Toyota vehicles into its fleet.

China has the second large robotaxi ecosystem. Baidu's Apollo Go was providing about 300,000 driverless rides per week by late 2025 and says its Wuhan operation has reached per-vehicle profitability, while Pony.ai and WeRide each passed 1,000 vehicles by the end of 2025. Pony.ai's fleet passed 1,700 vehicles by May 2026, and the company raised its year-end target to 3,500; its seventh-generation vehicles reached unit-economics breakeven in Guangzhou and Shenzhen, averaging 23 orders and 338 yuan of net revenue per vehicle per day in Shenzhen, although the company as a whole still reports losses. WeRide operates about 1,300 vehicles worldwide, plans to reach 2,600 by the end of 2026 using its purpose-built GXR robotaxi, and, like Pony.ai and Apollo Go, has expanded to the Middle East, Southeast Asia, and Europe.

Tesla launched its Robotaxi service in Austin in June 2025 with safety monitors on board, began unsupervised rides in early 2026, expanded to Dallas and Houston in April 2026, and widened its Austin service area to the full metro region in June 2026. Its unsupervised fleet has grown more slowly than Elon Musk's earlier predictions: crowdsourced tracking counted about 39 verified unsupervised vehicles in June 2026. Tesla began volume production of its two-seat Cybercab at Giga Texas in April 2026; the company says the design meets federal safety standards, which lets it self-certify instead of relying on a capped exemption, and it has also built a version with a steering wheel.

Other entrants are approaching commercial service. Amazon's Zoox, which builds a bidirectional vehicle with no steering wheel, served more than half a million riders for free before beginning paid service in Las Vegas on August 10, 2026, and it continues free rides in San Francisco while it seeks California approval. Wayve and Uber planned robotaxi trials in London under the UK's accelerated pilot program, and Nissan, Uber, and Wayve agreed in March 2026 to prepare a Tokyo pilot using the Nissan LEAF by late 2026, part of an Uber and Wayve plan covering more than ten cities. XPeng received road-testing approval for its GX robotaxi in Guangzhou and plans pilot passenger operations in the second half of 2026, with Amap, Alibaba's mapping service, as its first ecosystem partner.

Consumer Driver Assistance

Most people experience automotive AI through Level 2 driver assistance in cars they own. Tesla's FSD (Supervised) handles city and highway driving but requires constant driver attention, so it remains a Level 2 feature even though the same software family powers Tesla's driverless robotaxis. GM's Super Cruise offers hands-free driving on more than 750,000 miles of mapped roads in the U.S. and Canada and has logged more than a billion customer miles; GM plans to introduce eyes-off Level 3 highway driving on the 2028 Cadillac Escalade IQ using lidar, radar, cameras, and a new centralized computer, with turquoise lights signaling when the system is in control. Mercedes-Benz's Drive Pilot was the first certified Level 3 system in Germany and later gained approval in California and Nevada, and its December 2024 update raised its German speed limit to 95 km/h, with a goal of 130 km/h by the end of the decade.

Chinese automakers have made advanced assistance a standard feature rather than an expensive option. BYD announced in February 2025 that its "God's Eye" system would come on all its models sold in China, including cars priced under 100,000 yuan, while XPeng, Huawei-partnered brands, and Xiaomi compete with city navigation features and the newer VLA models described above. Nissan plans to bring Wayve's AI to its next-generation ProPILOT system from 2027. Because Level 2 systems keep the human responsible, camera-based driver monitoring that checks whether the driver is watching the road has become a central safety feature, and researchers continue to study how drivers' attention changes as these systems become more capable.

Autonomous Trucking and Logistics

Long-haul trucking is an attractive early market because highways are more structured than city streets and freight can move around the clock. Aurora began driverless commercial deliveries between Dallas and Houston in May 2025, the first such service with heavy-duty trucks on public roads, and later added lanes from Fort Worth to El Paso, El Paso to Phoenix, Fort Worth to Phoenix, and Dallas to Laredo. The roughly 1,000-mile Fort Worth to Phoenix route exceeds the federal hours-of-service limit for a single human driver, and Aurora reported more than 250,000 driverless miles as of January 2026 with no collisions attributed to the Aurora Driver. Weather limited its Texas driverless operations about 40 percent of the time in 2025, so a key software update added rain, fog, and high-wind capability, and the company plans to have more than 200 driverless trucks operating by the end of 2026.

Competitors are close behind. Kodiak AI began regular autonomous freight runs between Dallas and Houston with Roehl Transport and plans a driverless launch by the end of 2026, working with Bosch on production-grade hardware, and the carrier Hirschbach signed a non-binding agreement to deploy up to 500 Aurora-powered trucks. Autonomy is reshaping the broader freight supply chain, from yard tractors and warehouse robots to middle-mile delivery vans, but it also raises concerns among truck drivers about job security and among the public about how driverless trucks will handle severe weather and unusual road conditions.

Rail, Maritime, Aviation, and Traffic Systems

AI is also improving other modes of transport. Railways use machine learning for predictive maintenance of track and rolling stock, for scheduling, and for safety, including acoustic detectors in India that warn train operators when elephants are near the tracks. Shipping companies such as Stena Line have worked with Hitachi to cut fuel use through AI-based voyage optimization, and drones and aerospace systems apply many of the same perception and planning techniques used in cars. On city streets, AI-based traffic management uses data from connected vehicles, phones, and roadside IoT sensors to retime signals and reduce delays, and vehicle-to-everything (V2X) links over telecommunications networks let cars and infrastructure share warnings; the Traffic Prediction section below describes several of these systems in detail.

Recent Developments

The biggest shift in 2025 and 2026 has been from pilots to businesses. Waymo's ridership doubled in under a year, Chinese operators reported vehicle-level profitability in several cities, and falling hardware costs made larger fleets financially plausible, although no major robotaxi operator has yet reported company-wide profits from driverless service, and Pony.ai still reported a net loss of about $45 million in the second quarter of 2026. Investors are watching whether Waymo can reach its target of one million weekly rides by the end of 2026, whether Tesla can scale its unsupervised fleet and Cybercab production, and whether Zoox's exemption becomes a template for other purpose-built vehicles.

Research is converging on large, general models. Reasoning VLA models such as NVIDIA's Alpamayo family, whose models passed 500,000 downloads on Hugging Face by August 2026, and XPeng's VLA 2.0 aim to handle rare situations by "thinking through" them, while open releases of models, simulators, and datasets are lowering the barrier for universities and smaller companies. World models such as GAIA-3 are being positioned as evaluation tools rather than just video generators, and the UK's DriveSafeSim project is testing whether regulators can accept simulation evidence from them. An open question is whether language-based reasoning measurably improves safety or mainly improves explainability, and how much extra computing latency it costs.

The period also brought reminders of the field's difficulties. GM shut down its Cruise robotaxi business in late 2024 after a 2023 crash in San Francisco led regulators to suspend its permits, and folded the technology into its personal-vehicle program. Chinese robotaxi operators suspended some services in the UAE in March 2026 because of regional tensions, showing how geopolitics can disrupt international expansion. In factories, Ford's experience of adding human inspectors alongside expanding AI inspection and labor unions' objections to humanoid robots at Hyundai show that automation in manufacturing, like automation on the road, is advancing with humans still firmly in the loop.

Forward Thinking

The talks in this section range from technical keynotes to public-facing explainers about where driverless technology is heading. Andrej Karpathy's CVPR 2021 talk explains how Tesla trained camera-only perception on data from its customer fleet, and Nan Li's Autonomy Talks session presents game-theoretic methods that help an autonomous vehicle interact safely with human drivers. An SAE International session describes how Penske and TuSimple moved freight with Level 4 trucks, an early milestone for the autonomous trucking efforts described above, although TuSimple later exited the U.S. trucking market. TED talks by Chris Urmson, who led Google's self-driving program before co-founding Aurora, and by Wanis Kabbaj, along with reports from TVO and CBC News, explore how a driverless car sees the road and how autonomous vehicles could reshape cities.

Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube

MIT Series

Lex Fridman's MIT course 6.S094, Deep Learning for Self-Driving Cars, was published openly online, and its lectures remain a clear introduction to the field's foundations. The videos below cover deep learning basics, the course introduction, computer vision for driving, a 2019 review of the state of the art, an overview of the 2018 course, and a guest lecture by Waymo's Drago Anguelov. Some specifics are now dated, since end-to-end and foundation-model approaches have advanced quickly since 2019, but the core ideas of perception, planning, and reinforcement learning for driving still apply.

Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube

Traffic Prediction

Traffic prediction estimates future speeds, volumes, and travel times on a road network, from a few minutes to several hours ahead. Navigation apps use it to compute estimated times of arrival (ETAs) and choose routes, cities use it to time traffic signals and manage incidents, and freight and ride-hailing companies use it to plan fleets and set prices. Early methods relied on statistical time-series models and Kalman filters, while modern systems use long short-term memory networks and graph neural networks that treat the road network as a graph, so congestion on one street can inform predictions for its neighbors. Spatio-temporal architectures such as DCRNN, STGCN, and Graph WaveNet, along with the Spatial-Temporal Dynamic Network, combine these spatial links with patterns over time, and they often add weather, events, and incident reports as extra forecasting inputs.

The best-known production example comes from Google. In 2020, DeepMind and the Google Maps team described a graph neural network that divides roads into "supersegments" of adjacent segments with shared traffic and predicts travel time for each one, improving real-time ETA accuracy by up to 50 percent in cities such as Berlin, Jakarta, São Paulo, Sydney, Tokyo, and Washington, D.C., with the largest gain (51 percent) in Taichung. The team found that graph neural networks were sensitive to the order and variety of training examples, so making training stable was as important as the architecture itself. Google's Project Green Light applies similar data to traffic signals: it models each intersection from Google Maps driving trends and recommends timing changes to city engineers, and tests at about 70 intersections in 12 cities in 2022 and 2023 showed up to 30 percent fewer stops and up to 10 percent lower emissions, although the emissions figures are modeled estimates rather than direct measurements.

The videos below show the field's progression. Pawel Gora's PyData talk introduces TensorTraffic, a machine learning approach to traffic prediction; Ramya Raghavendra's Databricks session shows how weather data improves forecasts; an HPCC Systems tech talk applies deep sequence learning to traffic data; and a talk by Tanwi Mallick and Prasanna Balaprakash of Argonne National Laboratory presents traffic forecasting as a large-scale scientific computing use case, connecting this work to smart city planning. Taken together, they trace the field's shift from classic time-series tools toward deep sequence models and graph-based methods that learn from the structure of the road network itself.

Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube

Car Price Prediction

Forecasting

Car price prediction is a classic supervised regression problem: estimating a vehicle's market value from its make, model, trim, year, mileage, condition, options, location, and accident history. Dealers, online marketplaces, lenders, insurers, and leasing companies use these models to price inventory, appraise trade-ins, set residual values for leases, and flag listings whose prices look suspicious, an application of anomaly detection and fraud detection. Common methods range from linear regression baselines to random forests and gradient boosting libraries such as XGBoost and LightGBM, which usually perform best on this kind of tabular data, while neural networks can add value when photos or free-text descriptions are included.

Most of the work in a real pricing model lies in the data. Data preprocessing cleans scraped listings, encodes categories such as model names, and often predicts the logarithm of price so that errors are proportional, while feature engineering adds signals such as vehicle age and miles driven per year; careful cross-validation guards against the overfitting that comes from memorizing rare trims. Prices also drift over time, as shown by the sharp rise in used-car values during the 2021 to 2022 chip shortage and by the rapid depreciation of some electric vehicles after new-car price cuts, so models must be retrained regularly and should account for battery health in used EVs. Because appraisals affect real customers, many companies pair their models with explainability tools that show which features drove a given estimate. The videos below walk through complete projects, including an end-to-end Kaggle-style project by Krish Naik, a web scraper paired with a learning model, a multiple-regression classroom activity, and a linear regression implementation.

Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube

Surround Vision

The Chevy Equinox's HD Surround Vision feature uses multiple cameras to create a digital overhead view of the vehicle's surroundings:

  • Cameras - The Equinox uses four cameras to create the overhead view: one in the front grille, two on the side mirrors, and one in the rear.
  • Image processing - The cameras feed video signals to an image-processing program in the vehicle's computer. The program analyzes and combines the images to create a synthetic, top-down view.
  • Display - The overhead view is displayed on the vehicle's infotainment system. The display may also show the front or rear camera view in addition to the overhead view, depending on the vehicle's gear and speed.

The HD Surround Vision feature is useful for parking, navigating tight spaces, and being aware of hazards. It works at low speeds, similar to the vehicle's front and rear park assist. Behind the display, each camera uses a wide-angle fisheye lens, and the software first corrects the lens distortion using factory calibration, then projects each image onto an assumed flat ground plane and blends the overlapping edges into a single seamless picture. Because the method assumes everything lies on the ground, tall objects such as posts or people can look stretched near the seams, which is why many systems also offer 3D views that let the driver rotate the virtual camera around the car. Chevrolet's own explainer video below shows how the camera images are combined.

Autonomous vehicles use the same multi-camera idea, but for the computer rather than the driver. Instead of stitching pictures, a neural network learns a bird's-eye-view (BEV) representation of the scene directly from all cameras at once; the influential BEVFormer model (ECCV 2022) uses Transformer spatial cross-attention to gather features from each camera view and temporal self-attention to fuse information from previous moments, and it matched the accuracy of lidar-based methods on the nuScenes benchmark at the time. Tesla's "occupancy networks" go a step further and predict which cells of a 3D volume around the car are occupied, which helps the vehicle handle irregular objects that do not fit predefined categories. These learned bird's-eye views have become the common format for sensor fusion and planning in modern driving stacks.

Load video
YouTube

Manufacturing

AI is changing how vehicles are built as much as how they drive. Automakers use computer vision to inspect welds, paint, and assemblies; anomaly detection and predictive maintenance to keep presses and robots running; digital twins built on platforms such as NVIDIA Omniverse to plan and validate factories before equipment is installed; and, increasingly, embodied AI in humanoid and mobile robots that move parts around the plant. Large single-piece aluminum castings, often called gigacasting or hyper die-casting, replace dozens of stamped and welded parts, which simplifies assembly but makes hidden internal defects more costly and has driven demand for AI-assisted X-ray and CT inspection. The same technology used for self-driving (cameras, neural networks, simulation, and specialized chips) is now being reused for factory robots, a convergence that companies such as Tesla, XPeng, and Hyundai describe as "physical AI."

Results so far are real but uneven. AI inspection catches defects people miss and works without fatigue, yet Ford's experience shows that experienced inspectors still recognize patterns that vision systems overlook, so the most effective factories combine automated checks with human-in-the-loop review. Most published figures on humanoid-robot productivity come from the manufacturers and robot makers themselves, labor unions have raised concerns about job losses, and the business case for humanoids versus conventional industrial robots is still being tested. AI is also used upstream in the supply chain to forecast parts demand and manage disruptions such as the semiconductor shortage of 2021 and 2022. The entries below describe how individual manufacturers apply these techniques.

Xiaomi

Xiaomi Auto is Xiaomi's automotive business. The company announced its entry into electric vehicles in 2021 and launched the Xiaomi SU7 in March 2024. It followed with the YU7 SUV in June 2025, delivered more than 410,000 vehicles in 2025, and passed 600,000 cumulative deliveries by February 2026, while its EV and AI business posted its first full-year operating profit for 2025. Chief executive Lei Jun set a target of 550,000 deliveries for 2026, although price cuts and rising component costs pushed the unit back to a quarterly loss early in the year.

Xiaomi approaches vehicle manufacturing as a combination of mechanical engineering, electrical engineering, software development, AI-enabled perception, and intelligent production. Its vehicles compete in China's highly competitive automotive market, particularly against established electric vehicle manufacturers and newer technology-oriented automotive companies. Its factory in Beijing's Yizhuang district sits on a 720,000-square-meter site and uses robots in all six core processes: pressing, casting, body assembly, painting, battery pack assembly, and final assembly, with a second plant added next to the first in 2025. A centerpiece is hyper die-casting: one large module that sits above the rear wheels would conventionally require 72 parts welded at 840 points, but Xiaomi casts it as a single piece of aluminum, cutting production time for that structure nearly in half.

X-Eye is Xiaomi's proprietary AI-based X-ray inspection system used in electric vehicle manufacturing. The technology combines industrial radiographic inspection with Machine Learning (ML) to identify potential defects inside manufactured components. Unlike conventional surface inspection, which examines visible characteristics of a component, X-ray inspection can reveal internal abnormalities that cannot be observed directly from the outside. This makes the technology particularly valuable for complex metal components and large integrated castings.

The robot-mounted X-Eye system checks hyper-castings for internal voids and cracks with a reported accuracy above 99 percent. According to Lei Jun, pairing X-ray imaging with an AI vision model lets the plant inspect a large die-cast part in about two seconds, roughly ten times faster and more than five times more accurate than manual checks. In late 2025 he also said humanoid robots would be widely used in Xiaomi's factories within five years, and the company has been hiring robotics specialists, including engineers focused on dexterous robotic hands. On the vehicle side, Xiaomi's Hyper Autonomous Driving (HAD) assistance system brings the same perception expertise into its cars.

Load video
YouTube

Xiaomi's official "Xiaomi EV Factory" video, about three minutes long, tours the highly automated Beijing plant described above. It is a short promotional piece rather than a technical lecture, but it gives a clear sense of the level of automation across the plant's production stages. It pairs well with the hyper die-casting and X-Eye discussion in this section.

BMW Group

BMW organizes its production strategy under the name iFACTORY, which emphasizes AI, digitalization, and sustainable energy. Its in-house AIQX (Artificial Intelligence Quality Next) platform uses cameras and acoustic and other sensors along the line to detect defects in real time and give immediate feedback to workers, supporting a zero-defect approach with 100 percent end-of-line monitoring. The plant in Debrecen, Hungary, which began series production of the electric BMW iX3 (the first Neue Klasse model) in late October 2025, was planned and validated entirely in a virtual factory built on NVIDIA Omniverse, where it celebrated a "virtual start of production" in March 2023, and BMW is rolling the same approach out to plants such as Munich. BMW has also said it aims to use its virtual factory models to generate synthetic data for training quality-inspection and transport-robot AI before the physical environment exists.

BMW is one of the most active automakers in testing humanoid robots. Figure AI's Figure 02 robots loaded sheet-metal parts for welding at the Spartanburg, South Carolina, plant during a trial of about 11 months, which Figure says supported production of more than 30,000 BMW X3 vehicles, and BMW has since moved on to evaluating the newer Figure 03 for material logistics and parts supply. Critics note that only a small number of robots took part, so the trial demonstrates feasibility more than scale. In parallel, BMW's Leipzig plant began testing the wheeled AEON humanoid from Hexagon Robotics in 2026, letting the company compare two different humanoid designs under real production conditions.

BYD

BYD, one of the world's largest makers of electric and plug-in hybrid vehicles, is highly vertically integrated: it produces its own batteries, power electronics, and many semiconductors, which helps it add advanced electronics to low-cost cars. In February 2025 it announced that its "God's Eye" driver assistance system would come on its entire lineup sold in China at no extra cost, covering 21 models including cars priced under 100,000 yuan. The system comes in three tiers: God's Eye A uses three lidar sensors and the 600-TOPS DiPilot 600 computer for Yangwang luxury models, God's Eye B uses a single lidar and the DiPilot 300 for Denza, Fang Cheng Bao, and top BYD models, and God's Eye C relies on a three-camera module and the DiPilot 100 for mainstream cars such as the Seagull and Dolphin. BYD says its fleet supplied about 72 million kilometers of assisted-driving training data per day in 2024, and it has offered to take responsibility for incidents during its automated parking feature, an example of how manufacturing scale turns into data scale.

Ford

Ford's main manufacturing AI is the Mobile AI Vision System (MAIVS), developed with IBM's Maximo Visual Inspection technology and deployed across 17 North American plants. It uses ordinary smartphone cameras mounted at inspection stations to photograph each assembly step and flag errors before they move down the line, and it has reportedly performed about 150 million inspections and flagged 400,000 quality issues. For the launch of the 2026 Expedition at its Kentucky Truck Plant, Ford added 1,200 inspections and six times as many AI-powered inspection tools as for the previous model, but it also hired 203 new human inspectors; Bloomberg reported that Ford has been rehiring inspectors after AI systems fell short of expectations, a reminder that automated and human inspection work best together.

Ford's next generation of electric vehicles changes the assembly process itself. The Ford Universal EV Platform and Universal EV Production System, announced in August 2025, replace the traditional single assembly line at the Louisville Assembly Plant with an "assembly tree" in which three large subassemblies are built in parallel and then joined. Ford says the design uses 20 percent fewer parts, 25 percent fewer fasteners, and 40 percent fewer workstations, with assembly about 15 percent faster, and the first vehicle, a midsize electric pickup, will use its own onboard computers to verify key assembly connections during production so that the vehicle effectively diagnoses its own assembly.

General Motors

GM expanded its collaboration with NVIDIA in March 2025 to bring AI into factories, vehicle design, and in-car systems. In manufacturing, GM is building digital twins of assembly lines with NVIDIA Omniverse and Cosmos so it can simulate and optimize production before making physical changes, and it is training robots for material handling, welding, and quality inspection with NVIDIA AI. After shutting down its Cruise robotaxi unit in late 2024, GM redirected that autonomy expertise into personal vehicles: its next-generation centralized computing platform, debuting on the 2028 Cadillac Escalade IQ, will run both eyes-off highway driving and conversational AI features. GM has also announced integration of Google's Gemini models for conversational AI in its vehicles.

Hyundai Motor Group

Hyundai Motor Group bought Boston Dynamics in 2021 and is the automaker most directly invested in making humanoid robots. At CES 2026 it unveiled a production version of the electric Atlas humanoid, and it plans to deploy Atlas at Hyundai Motor Group Metaplant America in Georgia starting in 2028, beginning with physically demanding tasks and material transport. The group has said it intends to build capacity to produce 30,000 robots a year by 2028, and it already uses Boston Dynamics' Spot quadruped robots for plant inspection. The plan has drawn strong objections from Hyundai's South Korean labor union, illustrating the social tension around factory automation. Hyundai also supplies Ioniq 5 vehicles for Waymo's robotaxi fleet, linking its manufacturing to the robotaxi market.

Mercedes-Benz

Mercedes-Benz has invested in Apptronik, the Austin, Texas, maker of the Apollo humanoid robot, and is testing Apollo in several of its plants, including sites in Germany and Hungary. The robots focus on logistics rather than vehicle assembly, carrying components to production lines and handling repetitive internal transport tasks, which lets Mercedes evaluate performance before any wider rollout. On the vehicle side, Mercedes was the first automaker to certify a Level 3 system in Germany and the United States and says lidar and centimeter-level positioning are essential to it; its new CLA uses NVIDIA's driving software, and NVIDIA has said Alpamayo-based reasoning features will roll out starting with the CLA in the United States.

Tesla

Tesla has pushed several manufacturing ideas that other automakers now copy. Its Giga Press machines cast large sections of the vehicle body in one piece, and its "unboxed" process builds major sections of a vehicle in parallel and joins them late, which Tesla says reduces factory floor space and cost. The Cybercab is the first vehicle built this way: the first unit came off the Giga Texas line in February 2026, volume production began in April 2026, and Elon Musk warned that output would start slowly before ramping. Tesla also ended production of the Model S and Model X in 2026 to free space at its Fremont factory for its Optimus humanoid robot.

Optimus illustrates both the promise and the uncertainty of humanoid robots. Tesla has used Optimus units internally for tasks such as handling battery cells and parts, and its third-generation design features hands with 22 degrees of freedom; Musk has described goals of producing a million robots a year at Fremont and far more at a planned Texas facility. Independent observers treat those targets as aspirational because Tesla has not published verified production volumes and, as of mid-2026, had not sold Optimus to outside customers. The robot reuses Tesla's vision-based neural network approach from FSD, making it one of the clearest examples of driving AI being repurposed for factory work.

Toyota

Toyota's software and automation work is led by Woven by Toyota, which develops the Arene vehicle software platform and operates Woven City, a test community for mobility technology near Mount Fuji in Shizuoka Prefecture that opened in 2025. In April 2025 Toyota and Waymo agreed to explore a partnership to develop a new autonomous vehicle platform and to bring Waymo's technology to Toyota's personally owned vehicles, with Toyota vehicles potentially joining Waymo's fleet; the agreement was preliminary, and its scope was left to evolve. Toyota also works with Pony.ai, whose seventh-generation robotaxi includes a model built on a Toyota platform, giving the world's largest automaker exposure to both U.S. and Chinese autonomy approaches.

XPeng

XPeng designs its own Turing AI chip, which entered mass production in the third quarter of 2025 and had shipped more than 200,000 units by early 2026, with a goal of about one million shipments that year; Volkswagen is the first external customer for both the chip and XPeng's VLA 2.0 driving model. The company's GX robotaxi carries four Turing chips delivering up to 3,000 TOPS and uses cameras without lidar or high-definition maps. XPeng applies the same "physical AI" stack to its second-generation IRON humanoid robot, which uses three Turing chips, a solid-state battery, and dexterous hands, and which it targets for mass production by the end of 2026, with steelmaker Baosteel named as a partner for industrial inspection work.

Additional Viewing

Load video
YouTube

In this episode of The TWIML AI Podcast (about 69 minutes), host Sam Charrington talks with Drago Anguelov, head of research at Waymo, about the new foundation model that powers the Waymo Driver. The conversation complements the "End-to-End Models, Vision-Language-Action Models, and World Models" section by discussing how a company with a mature modular stack is adopting large learned models, and it adds context to the "Robotaxis" section on what it takes to scale driverless service.

Load video
YouTube

Wayve's chief scientist Jamie Shotton presents the company's end-to-end autonomy stack in this 29-minute talk at an ICCV 2025 workshop on foundation models for autonomous driving. It gives a first-hand view of the end-to-end learning philosophy described in the "End-to-End Models, Vision-Language-Action Models, and World Models" section and of the technology behind the Wayve robotaxi plans covered in the "Robotaxis" section.

Load video
YouTube

This NVIDIA Omniverse livestream (about 59 minutes) introduces Alpamayo 2 Super, NVIDIA's open reasoning vision-language-action model for robotaxis. It is the best companion to the discussion of VLA models, chain-of-thought explanations, and open simulation tools in the "End-to-End Models, Vision-Language-Action Models, and World Models" and "Recent Developments" sections.

Load video
YouTube

Carnegie Mellon University professor Phil Koopman, a leading independent voice on AV safety and a principal author of the UL 4600 safety standard, explains in this 43-minute lecture what it means for a self-driving vehicle to be safe once the human driver is removed. It provides the independent, critical perspective behind the "Safety, Evaluation, and Validation" section on how AV safety should be argued and measured.

Load video
YouTube

DeepFindr's 14-minute tutorial is a friendly introduction to temporal graph neural networks with a traffic-forecasting example. Although shorter than the other selections, it clearly illustrates how a road network becomes a graph and how spatial and temporal information are combined, which is the core idea in the "Traffic Prediction" section.

Load video
YouTube

PCMag's seven-minute condensed version of Hyundai Motor Group's CES 2026 presentation shows the production version of Boston Dynamics' Atlas humanoid and Hyundai's plans to put it to work in its factories. It is a concise visual companion to the "Hyundai Motor Group" entry and the discussion of humanoid robots in the "Manufacturing" section.

References

Retrieved from "https://primo.ai/w/index.php?title=Vehicles&oldid=55459"