Baidu Trains Robotaxis in Virtual City with Nonexistent Cyclists and Dogs

At a glance

Baidu has built Apollo World, a virtual city simulator, to practice rare or dangerous road situations that are hard to test on real roads. It creates hard-to-predict scenarios, such as cyclists or dogs suddenly running into the street, to test its robotaxis. The system is already reportedly used for Baidu's Apollo Go service.

Key points

1. 1

   The virtual cyclists and dogs on the virtual roads are driven by their own AI, so how they react depends on how the robotaxi behaves.

2. 2

   The system generates camera and lidar (a sensor that measures distance with lasers) data in real time, producing images similar to those from real sensors.

3. 3

   Baidu has not disclosed numerically how closely the virtual cyclists resemble real people, and there has been no verification by an outside organization.

4. 4

   A weakness is that when run for long periods, the model tends to drift away from reality.

Why it matters

It shows how driverless robotaxis are checked for safety before they go onto real roads.