The hype and media coverage may have died down a bit but many car makers still believe that autonomous is the future of driving. What many won’t admit, however, is how difficult it really is to train those self-driving systems. Most of them require going through known roads multiple times using complicated data from pre-scanned areas. MIT researchers, however, are developing an AI that can go through new territory with just a GPS-based map and some human intelligence.
We might still be far away from an AI that can learn with the same intuition as humans. Most AI really learn by going through massive amounts of data at high-speed but that has the disadvantage of knowing the data set beforehand. When driving through unfamiliar roads, that AI might start to fail.
The AI developed by Alexander Amini and Daniela Rus from MIT do study data sets but not of the roads themselves but how humans drive on roads, taking cues for signs, structures, and maps. An automated Toyota Prius is then set on a preplanned route in a different area and manages to make decisions almost exactly like a human.
There are definitely many benefits to this end-to-end navigation system. In theory, one can simply download a map into the car and it would be able to go from point A to point B without prior training in that area. It also means that the car won’t need expensive sensors like LIDARs or advanced mapping solutions.
And even when there is a discrepancy between the downloaded map and what the car sees through its cameras, the self-driving AI will still be able to make decisions based on previously learned driving methods. But perhaps better than most humans, it will also take the safest paths or make the safest decisions when it gets lost.