The latest fatality involving a self-driving Uber car highlights the reality that the technology is still not all set for ubiquitous adoption. The fact is that not many places are there where prevalent self-driving cars could actually drive in a reliable way. The fleets of companies such as Google are only tested in major cities, where they have spent considerable time in painstakingly labeling the accurate 3D positions of stop signs, curbs, and lanes.
“The cars use these maps to know where they are and what to do in the presence of new obstacles like pedestrians and other cars,” stated Daniela Rus, director of MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL). “The need for dense 3D maps limits the places where self-driving cars can operate.”
In fact, if one lives along the millions of miles of unreliably marked, unlit, or unpaved U.S. roads, then they are not lucky enough. Streets such as those are usually even more complicated to map and receive very less traffic; hence, companies are not incentivized to create 3D maps for them in the near future. Self-driving cars are simply not ready for huge swaths of America, from California’s Mojave Desert to Vermont’s White Mountains.
One solution for overcoming this is to develop systems that are sophisticated enough to steer without the help of maps. As a crucial first step, Rus and his team from CSAIL have created a framework called MapLite, which enables self-driving cars to drive on roads they have traveled earlier without the need for 3D maps.
MapLite integrates simple GPS data found on Google Maps with a sequence of sensors that observe the road conditions. Together, these two elements enabled the researchers to autonomously drive on various unpaved country roads in Devens, Massachusetts, and reliably detect the road over 100 feet in advance. (As part of a partnership with the Toyota Research Institute, the team used a Toyota Prius in which they fitted a range of IMU and LIDAR sensors.)
The reason this kind of ‘map-less’ approach hasn’t really been done before is because it is generally much harder to reach the same accuracy and reliability as with detailed maps. A system like this that can navigate just with on-board sensors shows the potential of self-driving cars being able to actually handle roads beyond the small number that tech companies have mapped.
Teddy Ort, CSAIL Graduate Student
Ort, Rus, and Ph.D. graduate Liam Paull, who is at present an assistant professor at the University of Montreal, are the co-authors of the paper, which will be presented at the International Conference on Robotics and Automation (ICRA) in Brisbane, Australia, in May 2018.
Despite all the advancements made with respect to self-driving cars, their navigation skills are still short of those of humans. Considering the way people get around on their own: while trying to reach a particular location, people probably feed an address into their phone and then consult it now and then along the way, such as while approaching highway exits or intersections.
On the other hand, if one were to travel the world similar to the majority of the self-driving cars, they would essentially be looking at the phone the entire time they walk. Prevalent systems are still more reliant on maps and only use vision algorithms and sensors to avoid dynamic objects such as other cars and pedestrians.
However, MapLite uses sensors for all facets of navigation and relies on GPS data only for acquiring a rough estimate of the location of the car. Initially, the system sets a final destination and also a “local navigation goal,” as per the researchers, which must be within the view of the car. Then, the perception sensors in the system produce a path to reach that point, making use of LIDAR for estimating the location of the road’s edges. MapLite has the ability to perform this without any physical road markings by making basic assumptions related to the way the road will be comparatively more flat than the adjacent areas.
“Our minimalist approach to mapping enables autonomous driving on country roads using local appearance and semantic features such as the presence of a parking spot or a side road,” stated Rus.
The researchers developed a system of “parameterized” models, indicating that they depict multiple situations that are quite similar. For instance, one model might be adequately broad to ascertain the course of actions at intersections, or on a particular type of road.
MapLite is different from other map-less driving strategies that are more dependent on machine learning in that it is trained on data from one set of roads and is then tested on the others.
At the end of the day, we want to be able to ask the car questions like ‘how many roads are merging at this intersection'? By using modeling techniques, if the system doesn’t work or is involved in an accident, we can better understand why.
Teddy Ort, CSAIL Graduate Student
Yet, MapLite has some limiting factors. For instance, it is not yet adequately dependable for mountain roads because it does not take into account drastic changes in elevation. The following move of the researchers is to increase the type of roads that can be handled by the vehicle. Eventually, their aim is to make their system achieve comparable levels of reliability and performance as mapped systems, however with a considerably broader range.
“I imagine that the self-driving cars of the future will always make some use of 3D maps in urban areas,” stated Ort. “But when called upon to take a trip off the beaten path, these vehicles will need to be as good as humans at driving on unfamiliar roads they have never seen before. We hope our work is a step in that direction.”
The National Science Foundation and the Toyota Research Initiative partially supported this study.