Artificial intelligence and EVs

Just two decades ago, state-of-the-art autonomous vehicles were confined to test tracks, and whiledriverless cars they already employed a wide array of sensors, the software was limited to rule-based and physics-driven methods. These methods are nearly impossible to generalise to real-world complexity. Machine learning algorithms, on the other hand, can leverage vast amounts of data to optimise driving performance and safety under real-world complexity. Computational power (exploiting graphical processing units [GPUs]) and AI algorithms (e.g. image recognition, transformer architecture) has seen rapid advances over the past two decades, making it possible to train neural networks which can handle very complex tasks such as autonomous driving today.

Autonomous vehicles typically employ a variety of sensors, such as cameras, radars and lidars, providing a 360° view of the surrounding environment. AI‑powered perception allows for accurate processing of large amounts of sensor data. AI algorithms are used to detect objects, as well as to create a coherent model of the environment using all available sensor data, and to make predictions about where objects will move next. Compared to humans, this allows for a more exhaustive view of the vehicle’s surroundings, which has been shown effective at avoiding accidents in common driving situations. However, difficulties remain in handling rare or ambiguous situations, which are not represented in training datasets. These can include extreme weather, rare objects, unexpected behaviour by other road users, and sensor malfunction. Handling of such cases is among the main R&D efforts for autonomous vehicle developers.

End-to-end AI approaches to autonomous driving could reduce costs and accelerate scale-up but are currently limited by regulatory and safety concerns. By integrating functions from perception to planning within a single model, these approaches have the potential to simplify development and deliver performance gains. However, ensuring safe and predictable behaviour is difficult: rare or unseen scenarios not captured in the training data can lead to unpredictable or unsafe actions, making certification highly complex. Today’s robotaxis employ modular approaches in which the planning of driving actions is still mostly rule‑based and thus easier to certify. As such, end-to-end approaches are likely to be largely confined to Level 2+ systems over the coming years.

Traditional battery management systems translate the voltage, current and temperature of a battery into state estimations using physics-based models. In an approach using AI, a machine learning model first learns relationships between voltage, current, temperature and charge/discharge history from lab data. Then, by collecting data from cars on the road, the lab-trained model can be fine-tuned with real-world data. The improved models are then deployed to the fleet via OTA updates. More accurate battery state estimations increase battery life and driving range, benefiting not only car owners but also OEMs, who can refine warranty strategies.

AI techniques can also improve the co-ordination of EV charging by predicting grid loads, EV arrivals and departures, renewable generation and electricity prices. These predictions are typically used by aggregators##anchor1## to set charging schedulesoptimise charging schedules for individual EVs. Relying on third-party optimisation of charging schedules allows EV owners to indirectly engage in real-time pricing and ancillary markets and contribute to local grid balancing, reducing charging costs. The potential for benefits through AI-enhanced charging co-ordination increases further with the availability of vehicle-to-grid charging.

Modern vehicles already depend on hundreds of semiconductor chips, and autonomous vehicles need significantly more. The semiconductor supply chain is characterised by tight supply and high concentration. Just one company produces roughly 70% of the world’s pure‑play foundry chips – chips that are designed by one company and manufactured by an external, specialised manufacturer.

Strengthening partnerships between automakers, chip designers and manufacturers is becoming increasingly important to secure access to critical components. This was underscored by the severe disruptions caused by the global chip shortages during the Covid-19 pandemic, and more recently by supply constraints experienced by the European automotive industry in 2025 linked to the Nexperia crisis. These crises have drawn significant government attention. For example, the European Union adopted the European Chips Act in 2022, aiming to mobilise EUR 15 billion (around USD 16 billion) in public and private investments by 2030, encompassing research projects, training, technical expertise, pilot lines, and equity support to semiconductor start-ups.

The growth of data centres and AI adds further pressure on chip supply chains. As chip producers – particularly manufacturers of memory chips – tend to favour data‑centre customers because of greater profit margins, rapidly rising demand could result in supply disruptions for other sectors, including the automotive sector. As of October 2025, it was reported that average memory‑chip inventories had fallen to around 3 weeks of demand, down from 15 weeks a year earlier.

As the industry transitions toward SDVs and increasingly integrates autonomous‑driving functions, ensuring stringent cyber‑security controls and adopting high security standards will be vital. The risks associated with cyber threats mean that cybersecurity and software update management has become an integral component of vehicle design and operation.

Cyber‑attacks could enable unauthorised access to data collected through in‑vehicle microphones, cameras and sensors, or allow malicious actors to disable or remotely operate vehicles. For example, it has been demonstrated that security gaps in a 2020 Nissan LEAF could be exploited to remotely control steering and other critical vehicle functions. Further back, in 2015, researchers remotely hacked a Jeep Cherokee to gain access to steering, braking and acceleration controls. VicOne, a cybersecurity company, identified more than 500 cyber vulnerabilities in vehicles already on the market in 2024.

Vulnerabilities in connected-car services can allow remote access to vehicles: for example, flaws in one brand’s system allowed attackers to remotely unlock cars via the mobile network. More recently, researchers identified web-portal vulnerabilities affecting connected vehicles that could enable remote tracking or control of functions such as unlocking, honking or remote start using only basic vehicle information. Cyber risks extend to EV infrastructure: research has demonstrated wireless attacks capable of disrupting EV fast-charging sessions, with potential implications for vehicles and power grids.