Researchers at Stanford University are working on a new set of controls for autonomous cars that integrate prior driving experiences. Ideally, this data can “teach” the car how to react in more challenging driving conditions.
Currently, control systems for autonomous cars need access to information on how the tires should interact with the road and the conditions of that road. This information is used to determine how hard the car applies the brakes, accelerates or steers. However, reacting to these conditions appropriately means an autonomous car needs to have been exposed to these situations previously and must be able to recall this data in order to apply the proper response.
This poses a number of distinct challenges, which is why the Stanford team exposed Niki, their autonomous Volkswagen GTI, and Shelley, the school’s autonomous Audi TTS, to a number of conditions in venues ranging from race tracks to the Arctic Circle. They’re hoping algorithms can be created from these experiences to develop a more flexible, responsive control system.
This neural network, or type of artificially intelligent computing system, will combine data from past driving experiences with foundational knowledge provided by 200,000 physics-based trajectories. In simulated tests, the neural network system outperformed the physics-based system. It also performed comparably to the performance of amateur race car drivers in identical driving conditions.
The results have been encouraging, but the researchers stress that their neural network system does not perform well in conditions outside of the ones it has experienced. They say as autonomous cars generate additional data to train their network, the cars should be able to handle a broader range of conditions, making them safer and more reliable.