Overcoming Limitations of AI and Machine Learning in AVs

By Graham Jarvis

---

12th June 2020

Autonomous vehicles are dependent on historical and real-time data, without which artificial intelligence and machine learning would be impossible.

They aren’t plug-and-play because not all of the potential scenarios can possibly be predicted by the software developers, simulators or data modelers. Simulators, and to a degree connected and autonomous vehicles (CAVs) themselves, are also only as good as their algorithms and the data inputted into them.

Consequently, AI and machine learning in autonomous vehicles can be limited, so nobody should expect them to instantly be able to cope with every potential scenario. Their development has to be taken with a sense of caution to prevent unintended consequences from occurring. The limitations of CAVs aren’t just about the AI and machine learning technology. There is also a need to educate consumers about what they can and cannot do safely. Perhaps, for this reason, there will be, for quite some time, the need for a human driver to have the ability to take back control.

Enabling AI

“Enabling AI everywhere requires device makers and developers to deliver machine learning locally on billions, and ultimately trillions of devices,” adds Dipti Vachani, senior vice-president and general manager of automotive and IoT line of business at Arm. “With these additions to our AI platform, no device is left behind as on-device ML on the tiniest devices will be the new normal, unleashing the potential of AI securely across a vast range of life-changing applications.”

Meanwhile, Danny Shapiro, senior director automotive at Nvidia claims that it’s hard to say “whether there is one location that is ahead of others” in the connected and autonomous vehicle market. However, he says there is so much development work ongoing with pilot projects running and “production vehicles are all occurring simultaneously”. He is certain that “everything that moves will eventually have some level of automated or autonomous capability”.

He explains that the key is a software defined platform: “Automakers around the world are recognizing the need for a high performance, AI computer that is programmable.” Shapiro sees the future of the automotive industry as being about software-defined self-driving cars requiring vehicle manufacturers around the world to “partner with companies that have the AI compute and software expertise to enable these capabilities to effectively compete. OEMs that embrace this strategy will be instrumental in helping their respective countries’ lead the AV race”.

Limitations of AI

As for the limitations of AI and machine learning, he points out that software-defined self-driving cars require massive amounts of computational performance.  He explains: “In fact, the amount of compute and the complexity of the software required has been underestimated in the past.  By building a centralized, energy-efficient AI supercomputer into the vehicle, car manufacturers will be able to update the software over-the-air and implement new business models that will increase the value of the vehicle over time.”

He also believes that manufacturers must also now reinvent their traditional business models and “look to integrate a single architecture that is richly programmable and can scale, allowing OEMs to have an entire fleet, from high-end models to entry level, that can be upgraded over time with new automated and intelligent features”.

Autonomous vehicles don’t just need massive amounts of computational performance; they also deep neural network development because these are responsible for the exhaustive training required for self-driving.  He explains: “They must be able to handle everything, from situations they’re likely to encounter during daily trips, to the unusual ones they’ll hopefully never come across. Success is about ensuring they’re trained on the right data.  By applying AI to this process, it’s possible to cut down on the time and expense to train these neural networks, while also increasing their accuracy.”

Overcoming limitations

So, what should be done to overcome the limitations of AI and machine learning in autonomous vehicles? The first aspect is data. They require massive amounts of it to operate efficiently and effectively. Shapiro adds: “Managing and curating this data requires high-performance compute, as well as intelligent training methods.

Mitigating latency

There is also a need to mitigate latency to ensure that real-time decisions can be made by the vehicles. To a certain extent edge computing can play a role, allowing data analysis close to where the data is created by the sensors and by any infrastructure with which the vehicles interact. They will also need to learn from, and to be able to intelligently interact with, other vehicles.

However, when collated together the data will be voluminous and require timely big data analysis. AI and machine learning can be used to expedite the transmission of big data and offer more accurate big data analysis to allow CAVs to make better decisions, improving vehicle performance and safety.

Shapiro concludes that the future of the autonomous vehicle will nevertheless be software-defined. However, that future has, in his opinion, to start today: “Running a complex software stack that enables the most advanced features today.”