Up until now, most detection systems have been trained through machine learning, say the researchers from Viterbi School of Engineering in California and Arizona State University. The problem: The resulting algorithms are difficult to test, since no one fully understands how they make their predictions. Now, new testing software is to detect errors in the algorithms long before the car hits the road. Simply put, the software tests, for example, if the system understands that an object cannot simply appear out of nothing or disappear from one frame to the next. If it fails the test, there is probably an error in the algorithm.

Many companies are also striving to improve the ‘perception’ capabilities of self-driving cars. One such company is startup Artisense, based in Garching, near Munich, whose founders are aiming to enable machines to see like humans. The startup’s software creates a complete streetscape in a 3D cloud out of individual points, reports German online news Website heise.de. In the process, artificial intelligence facilitates 3D mapping and real-time localization on the map.