Comment: Can AI take drivers out of the loop?

If you were in Las Vegas for the CES Show earlier this year, the one bet the automotive sector is putting its money on is AI. From the keynote speeches through to the slick car maker reveals, it appears AI has come of age for the automotive industry as a whole.

Projects such as GM’s Dreamcatcher where AI-powered generative design offered double digit weight and strength gains has long since shown what AI could do for engineers, but now it’s being touted as a pathway to improving in-vehicle customer experience.

Anyone who has ever been frustrated by an intrusive lane assist system or a distracting HMI menu, will know that before any AI algorithms are deployed, human-centric training and testing are needed to create bias-free datasets and, ultimately, marketplace acceptance. Betting the house on leaving it all to AI could have unexpected consequences. Forward thinking OEMs have grasped this, deploying Driver-in-the-Loop (DIL) simulators to inject human engagement in the design and validation of vehicles, both early and often.

At the heart of relevant and applicable DIL simulation is the all-important, live interaction between humans and proposed technologies. When humans control a vehicle's trajectory, they function as the 'control system' for the 'plant' that is the vehicle. A human driver provides inputs – throttle, brake, steering, gear selection – and the vehicle responds accordingly, providing real-time physical ‘feedback’ to the human driver which informs corrective ‘control’ actions. But what about the growing number of cases when the car itself is in the driver’s seat?

While AI has made significant strides in autonomous piloting technologies, it still requires human-derived data to learn and improve. Research by McKinsey found that AI systems can achieve up to 15 per cent higher performance when trained with high-quality human feedback. This feedback loop is crucial for realistic insight and vital for optimising vehicle performance and engagement. It’s best served through either physical mileage accumulation on real roads or test tracks, or, increasingly - through DIL simulation.

DIL simulators are gaining ground for several reasons. Not only do they provide controlled, repeatable laboratory environments where AI algorithms can be trained and evaluated to enhance autonomous driving systems, but DIL simulation labs also provide an inherently safe space where edge cases – extremely important for AI training – can be explored without putting people or hardware at risk. And for any vehicle maker aiming to reduce resource consumption and hit carbon reduction targets, DIL simulators offer clear advantages over traditional testing approaches.

Ultimately, while AI can ostensibly cover thousands of scenarios, it still needs to be developed and trained with nuanced feedback from human drivers and occupants to refine systems for successful real-world deployments.

Take that earlier example of the mandatory introduction of Lane Keep Assist and Autonomous Emergency Braking in 2022. Does that intentional interruption of the human driving (control) tasks, work in every market or culture? Will the handwheel loading cause a human to respond too aggressively or even take their hands of the wheel? Could they be distracted or irritated by the interruption to an extent that disrupts their brand loyalty?

One might argue that nowadays the value of the human "in-the-loop" component is directly proportional to the level of intervention from the assistive component, i.e. more ADAS warrants more DIL simulation and assessment, rather than less. A study by the National Highway Traffic Safety Administration (NHTSA) found that incorporating human feedback in the development of autonomous vehicles can reduce the occurrence of critical safety events by 30 percent, ensuring a safer and more reliable deployment of these technologies.

As the capabilities of AI evolve, DIL simulation is and will continue to advance in parallel, providing full-vehicle and system-specific evaluation opportunities. The key here is recognising that human assessment and acceptance will remain at the centre of vehicle developments – and increasingly so. Particularly, as the latest engineering-class DIL simulators become ever more powerful and versatile, evolving into vital tools for so many vehicle attributes – not just ride, handling, HMI or active safety functions. As well as greater fidelity, more immersivity and reduced latency, there is a growing ability to integrate Hardware-in-the-Loop (HIL) interactions. From a camera to a full EV powertrain, everything can be integrated regardless of whether or not a simulator and hardware test bench are even located in the same facility. Those who are embracing DIL simulation are circumventing traditional build-and-test approaches and fully reaping the biggest rewards of improved product acceptance, reduced development cycles, lower costs and minimised environmental footprints.

Going back to Vegas, it’s quite easy to get swept away dreaming of big pay-outs. But when the chips are down – as they may seem to be in an increasingly competitive and tech-aggressive automotive space – gambling it all on AI isn’t the sort of bet vehicle manufacturers can afford to make.

Salman Safdar, subject matter expert at Ansible Motion