Researchers at theMassachusetts Institute of Technology have created a photorealistic simulation system to train driverless cars prior to their deployment on actual streets.
The Virtual Image Synthesis and Transformation for Autonomy (VISTA) system employs a small dataset assembled from the experiences of actual drivers, to synthesize a virtually infinite number of pathways the vehicle could take in the real world.
The controller is rewarded for the distance it travels without crashing, which requires it to learn how to safely reach a destination by itself, and navigate any situation it encounters.
Tests showed that a controller trained via such reinforcement learning in simulation could be safely deployed into a driverless car and navigate through previously unseen streets, the researchers said.
From MIT News
View Full Article
Abstracts Copyright © 2020 SmithBucklin, Washington, DC, USA