New MIT simulation system teaches autonomous autos real-world crash avoidance

Scientists at MIT have created a new simulation system that they invented to train autonomous autos with infinite steering possibilities. The goal of the simulation system is to help the autonomous autos learn to navigate a range of worst-case scenarios before they are let loose on real streets around the country and world. Control systems for autonomous autos rely heavily on real-world data sets of driving trajectories from human drivers.

From that data, the self-driving cars can learn to emulate safe steering controls in a variety of situations. Real-world data from hazardous "edge cases" such as nearly crashing or being forced off the road or in the other lanes is rare. Computer programs known as "simulation engines" aim to imitate the situations by rendering detailed virtual worlds to help train the autonomous auto controllers to recover from these potential situations.

The problem in these situations is that the learned control from the simulation is never transferred to reality on a full-scale autonomous vehicle. MIT researchers have addressed that problem using a photorealistic simulator called the Virtual Image Synthesis and Transformation for Autonomy. The system uses a small data set that's captured by humans driving on a road to synthesize a practically infinite number of viewpoints from trajectories that the vehicle could take the real world.

The controller is rewarded for the distance it travels without crashing, so it learns by itself how to reach a destination safely. As the controller works to reach its destination safely, it learns to navigate any situation it encounters, including regaining control after swerving between lanes or a recovery from a near-crash. In testing, the MIT simulator was able to be safely deployed onto a full-scale driverless car and navigate through previously unseen streets.

When the vehicle was positioned at off-road orientations that mimicked various near-crash situations, the controller was able to successfully recover the car back into a safe driving trajectory within a few seconds. MIT's work was performed in collaboration with the Toyota Research Institute.