Discover more about the topics and technologies to be discussed at this year's conference, via a series of exclusive interviews with a selection of our expert speakers
Dominique Seydel, research engineer at Fraunhofer ESK, discusses AI’s benefits, but explains why we need to acknowledge its limitations and use resilient software to work around them.
Which AI features are already mature and usable in off-highway vehicles today and what are the designated applications for the future?
Today’s applications utilizing AI mainly focus on perception of situations by receiving sensor data as input and then deriving driving decisions for the actuators. In terms of controllability and reliability, however, there is still some way to go before a fully production-ready solution exists. Therefore, the AI can only be used in separate steps of the data processing chain.
The perception of the environment from sensor data is one step where AI can play out its strengths. Concrete examples include the detection of road markings and obstacles in the path of the machine such as other vehicles, people and power cables. AI can be a powerful tool for sensor fusion, but beyond that, can also help to predict the movements of potential hazards. In the future, AI can therefore contribute to mastering complex driving situations like collective tasks and safely adapt to unlearned driving situations.
What are the current weaknesses of AI?
AI doesn’t always recognize objects from video data in the same way that humans do and, consequently, as we would expect. For example, a train may not be recognized by its shape, but instead by the typical characteristics of tracks. Since the results of the AI depend largely on the training data with which the neural network was trained, there would have to be an enormous variance in this training data. Only in this way can we be sure that variables like the weather, the time of day, the position of the sun and changed road markings do not change the result of the object recognition.
Today, we do not yet have sufficient training data sets that can train AI to reliably recognize all possible situations. Calculations indicate that billions of test kilometers from all parts of civilization would be needed to truly validate the AI. This is simply not feasible at reasonable cost. In addition, it cannot be guaranteed that all potential corner cases can be foreseen and are present in the training data.
The biggest disadvantage for the reliable use of an AI in safety-critical applications is therefore that we cannot understand how results are obtained in the AI to prove its safety. This proof is necessary in order to approve AI-based vehicles at all, given liability issues and because of the responsibility of vehicle manufacturers toward customers and other road users.
What measures do you propose to rein in and control AI, to make it failsafe?
Our approach to securing the results from AI is to consider it as not sufficiently trustworthy. Therefore, we extend the software architecture to include an additional path of sensor data processing designed solely to ensure safety. This safety path recognizes with sufficient reliability – but possibly less performance – conditions that may cause the AI to fail to recognize obstacles.
The aim of our safety approach is to reach a level of system maturity where the data from all the individual sensors is included in the processing when environmental conditions demand it. Of course, the challenge lies in restricting the AI to ensure safety, but at the same time not impeding its potential.
How long do you think it will be before human operators won’t be needed anymore?
There is a lot of activity in the market and a lot of different predictions have been made over time about when humans will become passengers. However, such predictions are still vague. Most indicate that Level 5 AVs will not be on public roads before 2030. Autonomous industrial vehicles, on the other hand, operate in more limited conditions. Here, the complexity also lies in its actual functionality and expected performance. Autonomous industrial systems shutting down to failsafe modes in the case of unforeseen events is not cost-effective, but for restricted and specific areas it will be possible to automate the driving functions earlier.
Dominque Seydel will give a presentation titled Resilient software architectures for autonomous systems as part of the Autonomous Industrial Vehicle Technology Conference. Click here to book your delegate pass, which gives you access to all four conferences.