Autonomous Vehicles: Hop to it

Audi already has the A8, the first level three car

Max Odendahl, CEO and co-founder of Silexica, discusses key multicore and technology challenges faced with autonomous vehicles and talks about how the industry can catch-up with the media hype through automation

A recent news story from Australia talked about how software is being developed for driverless cars to enable them to avoid kangaroos on the road. The animal is very unpredictable and can hop quickly in any direction, so General Motors has taken to creating a dummy kangaroo to try to forecast the movements and implement these into its automotive trials.

Volvo is also working on this problem and says it’s large animal detection system enables cars to avoid European animals such as deer and elk. However, their movements are more stable, the kangaroo is not so straight-forward.

As a result, there are a team of engineers in Australia gathering information on the movements of kangaroos.

This is a typical day in autonomous driving, each resulting in numerous media stories from every corner of the world. The marketing and public relations efforts in the automotive sector seem to have overtaken engineering reality and the public, without knowing the levels of autonomy from zero to five, are starting to believe that completely driverless level five cars will be available imminently from their local dealer. Before we get to that stage, engineers have many pressing problems to solve, including some even more difficult than kangaroos.

 

Multicore

The biggest challenge from the automotive industry is how is it possible to squeeze all the required power and processing ability into a tiny embedded computer that takes up as little space as possible?

At the moment, multi-board supercomputers take up the whole boot. They are heavy and produce a lot of heat, and are still one of the major obstacles the OEMs and tier-ones must address before having a free boot can become a reality. To change this, a real understanding of system behaviour in dynamic scenarios is needed and the correct platform chosen. This is easier said than done as the demand for computing power has never been as great. Nobody knows which platform will be able to deliver this power, and if it is found, nobody knows how to programme these untested platforms. This is all combined with the intense pressure to be the first to deliver the mass-produced vehicle the world seems to crave.

There is an assumption that boards such as the NVidia PX have the performance ability. These are indeed raising the bar in terms of processing power available. To get the absolute best out of these boards, it needs software automation to take the vast numbers of algorithms onto such heterogeneous systems.

The first step should be to analyse fully the code in a matter of hours through automation compared with months via manual processing. If this can be done, then the engineer is in a much better position to move forward with the project as it frees time to focus on other tasks.

Once there is a complete understanding of the code, the engineers can seek to optimise the distribution to the target platform. By simulating techniques and forecasting behaviour of the software application, they can get detailed feedback on the execution flow. With this information, they can really push the performance, power and memory constraints to increase performance.

Then they are ready to implement their software based on this detailed and precise information. It frees up space both physically and electronically.

In some cases, after analysing, optimising and looking to implement the software, the results may show that this is not the most efficient way to proceed. By having automated all of this, the engineer has saved a lot of time and money and can thus reduce risks to deliver the project.  

 

Deep learning

The second challenge is in terms of deep learning and neural networks. This area includes forecasting the unpredictable, the weather, construction works, animals and, yes, the kangaroo. Training a computer system to reduce usage of datasets so it can act faster and in smaller systems is crucial in the development of these vehicles.

There is a lot of research and development and major investment going into this area. The critical part is how to ensure it doesn’t fail after training. It has to be perfect. One early failure, for example, such as a brake applied too late could ruin the reputation of a driverless car before it has even begun, especially given the media watching every move. Take for example the self-driving Uber collision in Arizona earlier this year; it was global news.

Again, by software technologies automating these processes and algorithms, and simulating them relentlessly, it is the only way to make absolutely sure this works by optimising the performance of new platforms.

 

Legislation

Finally, and not a software topic but still something that needs to be addressed, is legislation. Who takes responsibility should something go wrong? The OEM? The tier one? The driver? It depends on the level of autonomy, but this potential legal minefield needs to be clarified and often the systems for this in some countries are not that quick.

The USA is pressing ahead and amending safety codes to speed up the roll out of autonomous vehicles and European countries should slowly follow. Audi already has the A8, the first level three car, which is packed with technology, cameras, radars and sensors enabling controls such as a traffic jam pilot. Still, this car is not approved yet in Germany. Different approvals will likely be needed from different countries and this can take time.

The strong safety case of erasing human error and reducing accidents is helping fuel the legal processes. However, on the other side, any accident or security breach with the world watching will also fuel the safety case against autonomous vehicles.

 

Conclusion

There really is no better time to be involved with software and electronics in the automotive sector. Engineers love solving problems and creating a part of the future, and this is a fascinating era of advancement with which to be involved. Autonomous cars are certainly coming; the big question is when. By automating processes between software and hardware, this can happen a lot sooner.

Max Odendahl is the CEO and co-founder of Silexica

www.silexica.com

Recommend to Colleagues: