Ces 2016 : Toyota Demoes Autonomous Driving Using Deep Learning


Toyota Demoes Autonomous Driving Using Deep LearningToyota Motor Corp showed a system that automatically learns driving skills through deep learning.
Toyota exhibited the system at 2016 International CES in collaboration with Preferred Networks Inc and Nippon Telegraph and Telephone (NTT) Corp. Toyota used eight model cars and showed how they had acquired driving skills after zero, two and four hours of learning.


The only information to be given to the system first is traveling route data. Data about sensors that recognize the surrounding environment is input to each vehicle's learning unit, which is connected to the accelerator, brakes, steering wheel, etc.

No meaning is given to sensor data or output information in advance. On the other hand, when the vehicle crashes into an obstacle, a penalty is given for reinforcement learning. As the vehicle keeps learning, it grows to be able to travel smoothly. The system collects the learning results from each vehicle, performs calculations and feeds them back to the learning unit of each vehicle to accelerate learning speed.

In the demonstration of the system at the trade show, the model cars were not equipped with sensors. Instead, cameras set at high places were used to take images of the cars, and mock data was created based on the images (in place of data collected by the sensors) and sent to the learning unit.