Supporters of autonomous vehicles point out the many advantages computer controlled vehicles offer; they don't get distracted, drink alcohol, go over the speed limit because they need to be somewhere. There are still challenges though, in particular how computers can handle the unexpected, such as pedestrians stepping into the road.
Pedestrian detection systems are available in commercial vehicles, but currently as an add-on aimed at allowing the car to respond to unexpected obstacles more quickly than the driver, so there is redundancy if either driver or computer fail to react. However, fully autonomous vehicles must be self-reliant, so the system must be able to operate reliably without a manual override.
Self-driving cars on our roads may well become a reality and it’s
happening much sooner than we think. Credit: STFC/Helen Towrie
Paolo Rech is an Associate Professor at the Federal University of Rio Grande do Sul (UFRGS), in Brazil. He has been studying the impact of neutrons on electronic devices, including working with companies like NVIDIA and AMD who produce Graphic Processing Units (GPUs) for the electronics industry. He has been using ChipIR (currently in commissioning) to test embedded GPUs for use in driverless cars against neutron strikes generated by cosmic rays entering the Earth's atmosphere from space. The aim is to establish whether off-the-shelf electronics meet the strict standards for use in the automotive industry, or whether it is necessary to improve algorithms' reliability with software fault-tolerance or even design something new, at significant extra cost to protect against these particles from space.
Prof Rech says, “In US and Europe there are standards for reliability of any electronic components, which state the error rate must be less than 10 errors in 109 hours of operation. Google have been testing driverless cars for about 3 years now, and their error rate is around 9000 in 109 hours of operation, so some way above this. However, looking at the reported accidents caused by human error, this rises to about 28000 errors in 109 hours. So even at present driverless cars could save a number of lives."
In his experiment on ChipIR, Prof Rech and his team used a simulation of a car driving around urban streets and responding to pedestrian incidents. They recorded how often the computer responded and how many errors it made. Paolo says, “Essentially, the computer spent hours watching a video, while we counted how many mistakes it made!"
They showed that in the system Prof Rech was testing, the number of neutron-induced errors was relatively small. Not only that, when errors were detected, 60% still elicited the correct response, i.e. all and only the pedestrians in the frames were detected. Prof Rech says, “We need to look at the results in a bit more detail, but so far it's looking good. There appear to be more false positives – that is, the car stopping when it doesn't need to – but the computer is already performing much better than its human counterparts."
ChipIR is only in the commissioning phase at the moment, so Prof Rech has only had a limited amount of time to study these devices, but he has many plans for the future! He says, “We were delighted to have access to ChipIR – even in the commissioning phase we have seen how the high flux benefits our experiments. ChipIR will form a key part of our research going forward, not just in testing pedestrian activation systems but in a whole host of electronic devices. We will be back before summer to test Artificial Neural Networks for self-driving cars and also High Performance Computing parallel devices designed by Intel and NVIDIA, to evaluate supercomputers reliability which is another key aspect of our research. Additionally, we will validate the software-based hardening strategies we are developing to improve parallel devices reliability for automotive applications."
Dr Chris Frost is the instrument scientist on ChipIR. He says, "It is really great have Paolo as one of the first users of ChipIr working on this important area of research. We designed ChipIr exactly for this kind of experiment so that people working on some of the most advanced electronics can understand how cosmic ray neutrons are affecting their systems and ensure that those systems do the job they are supposed to do in a resilient and reliable way, before they hit the market. At the end of the day the best outcome of this research is that no-one even notices the impact of cosmic rays, because this knowledge was built in at the design stage!"