MIT Researchers Create System That Trains Driverless Cars Before They Are on the Road

Vehicles learn to drive in real-world scenarios before hitting the road.
Fabienne Lang
MIT simulationMIT

Driverless cars are becoming more and more of a reality, and with safety measures, testing and training them becomes a priority. 

A team of researchers from MIT has created a simulation system to train driverless cars in real-world contexts, helping them learn how to navigate through worst-case scenarios.

Their findings were published in the journal IEEE Xplore.


Photorealistic simulation engine

Control systems that are currently used to train autonomous vehicles largely rely on real-world datasets from human drivers. However, real-world situations can be much more hazardous, involving near-crashes or being forced off of the road or into other lanes. 

So far there have been some computer programs that try and imitate these types of situations with virtual roads, however, the learned control from this type of simulation has never shown to transfer properly and directly to reality. 

Now, researchers from MIT have designed a photorealistic simulator, called Virtual Image Synthesis and Transformation for Autonomy (VISTA).

It uses a small dataset snapped by real drivers on roads, offering a huge array of different vantage points. The controller is rewarded for the amount of time spent on the road without having any issues, teaching itself how to maneuver the streets safely. This includes recovering from having to quickly swerve out of the way or recovering from near-crashes. 

In their tests, a controller that had successfully undergone the VISTA simulation was able to be safely set out in a completely driverless car and drive down previously unknown streets. The system was put through a number of test near-crashes and was able to fully recover and keep control of the car. 

Alexander Amini, a Ph.D. student in Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT stated "In our simulation, however, control systems can experience those situations, learn for themselves to recover from them, and remain robust when deployed onto vehicles in the real world."

The authors successfully drove 10,000 kilometers (6,213 miles) in simulation before applying their fully autonomous vehicle to the real world. As far as they are aware, the authors state this is the first time a controller was trained in this way

Amini said "That was surprising to us. Not only has the controller never been on a real car before, but it’s also never even seen the roads before and has no prior knowledge on how humans drive."

The researchers keep up their work, as they now include other factors into their system, which include rainy or sunny weather, night and day, etc. 

Add Interesting Engineering to your Google News feed.
Add Interesting Engineering to your Google News feed.
message circleSHOW COMMENT (1)chevron
Job Board