It’s a situation that every driver has experienced dozens of times. You’re in the left lane on the highway and instinctively foresee that a car in front is intending to move into the left lane for a passing maneuver and will shortly pull in front of your vehicle. People have a kind of sixth sense for such situations: based on the context and their years of experience in traffic, they conclude that the other driver could be about to cut them off.
Machines still struggle to draw such conclusions. That’s why Adaptive Cruise Control (ACC) currently only reacts when a car actually moves in front of the vehicle from the right or left. “Based on our experience in series development, we came up with an innovative idea: it would be much more convenient if the technology could anticipate the cut-in maneuver and create some space beforehand – just as any experienced human driver does. That’s the idea behind ‘cut-in-detection,’ which should make driving even more practical in the future,” explains Philipp Wustmann, Project Manager for Driver Assistance Systems.
Without artificial intelligence (AI), this idea is difficult to implement because the indications of cutting in can only be described to a limited extent by rules and traditional programming.
That’s why Porsche Engineering’s location in Cluj-Napoca, Romania, comes into play when it comes to cut-in detection: the software and function developers there work on new vehicle functions, but also innovative development tools.
“An important factor in our future success will be the development and use of AI methods and corresponding tools,” says Dirk Lappe, Managing Director of Porsche Engineering. “The innovative environment and data science expertise at the Cluj location will enable us to break new ground in vehicle development.” Some 160 employees are spread over four floors in a new building blessed with copious natural light—and numerous AI experts. New colleagues are added every month, and the location is set to grow to 210 staff by the middle of 2020.
New solutions soon in series production
The experts in Cluj-Napoca use the latest AI methods such as neural networks or reinforcement learning to develop new driver assistance systems, the autonomous vehicles of the future, or virtual tests of new vehicle functions. “Here we are developing new solutions and features that can go into series production in the near future,” says Marius Mihailovici, Managing Director of Porsche Engineering in Romania. “By working on exciting development projects and cooperating closely with universities, we are continuously expanding our portfolio and our capabilities.”
The cut-in detection function uses a neural network: people look at pictures of cars traveling on a highway and press a button as soon as they predict that a car in front will shortly cut in. “In the training phase, the neural network links this human input with the signals provided by the vehicle sensors,” explains Rares Barbantan, a software architect at Porsche Engineering in Cluj-Napoca. “So it learns to predict the cut-in from the sensor data.” With this knowledge, the ACC can react at an early stage to maintain the necessary distance from the vehicle cutting in. The developers in Cluj-Napoca have been working on the new system since late 2018. In many cases, cut-in detection does not require additional sensors—the main information used is the list of objects in the two nearest lanes that is already available for other assistance systems.
The AI experts in Romania also want to further improve adaptive cruise control itself. More precisely: they want to make it more practical and, in the future, able to adapt to the driving style of the vehicle owner. “In the first step, the development colleagues in Prague implemented the core function of adaptive cruise control using conventional methods. In the next step, it is now learning to optimize its behavior with AI in order to adapt in accordance with the preferences of the driver,” explains Tudor Ziman, who directs software development for new functions in Cluj-Napoca. This process involves the use of reinforcement learning: in the simulator and on the road, people evaluate how practical they find adaptive cruise control. From this feedback, the system learns to respond more like a human.
Please click here to view the full press release.