MIT study combines basic map and sensors for ‘mapless’ autonomous driving

Researchers at MIT, in collaboration with Toyota, have developed a system for self-driving cars that works without detailed maps of an area. They combine basic maps with information from sensors on the car.

In their research, the scientists state that self-driving cars are highly dependent on detailed and annotated maps, which indicate curves and road signs, for example. This becomes a problem once an autonomous car enters an area that has not been mapped precisely enough. Their self-developed system, which bears the name MapLite, should offer a solution for this. It uses basic Openstreetmap maps without much detail. The car then determines exactly where it is based on various sensors, including lidar and GPS. In this way, the car must be able to take a road that has not yet been visited.

Their approach to so-called mapless driving, of which there are several, would have the advantage of eliminating the need to keep detailed maps up to date. The researchers do mention that the Openstreetmap map dataset contains the applicable traffic rules for each road section. With the help of the lidar, in this case an HDL-64 from Velodyne, it is possible to see 35 meters in front of the car. According to the researchers, this should make it possible to achieve speeds above 107 km/h. If the data processing were to take place in parallel on a GPU, the speed could be increased even more.

One of the researchers, Teddy Ort, tells IEEE Spectrum that a global map suitable for their system would fit on a USB drive. A detailed map of a small town, on the other hand, would require a few gigabytes of space. One of the drawbacks of their approach would be road safety checks. This would be less difficult on roads where cars have already driven over them. Another limitation is that the system is not yet capable of navigating mountain roads as it cannot cope with sharp ascents and descents.