April 03, 2018

What it Takes to Drive Autonomously on Chinese Roads

Yiming Liu presenting at GTC 2018

Last week at NVIDIA’s annual GPU Technology Conference (GTC) in San Jose, Pony.ai’s Infrastructure lead, Yiming Liu, shared anecdotes from our experience driving and operating in China.

The uniquely challenging urban environment in China has truly pushed us to new heights in our technology. When we first began testing in Guangzhou roughly four months ago, we noticed that road behavior in urban China is quite different than that of the US. The challenging environment, based on our observation, is largely driven by two key behavior changes:

  1. Traffic rules are constantly broken, sometimes in a more extreme fashion than in the US
  2. Highly unpredictable driving and pedestrian behavior

The type of behavior shown in the video above encapsulates scenarios we encounter daily. In fact, this footage is taken straight from the cameras we have mounted on our vehicles during a daily run–both scenarios occurred within 5 minutes of one another.

As you can imagine, this environment posed even more stringent requirements on our technology. In order to handle these scenarios safely and effectively, our entire system had to be even more efficient in real-time processing, decision making, and response. The margin of error for real-time object classification and speed detection, though seemingly impossible, became even slimmer.

To address these challenges, we employed an adaptive model of deep sensor fusion, one that intelligently relies more heavily on certain sensors depending on the driving scenario. A more powerful on-vehicle computing system also allowed us to support more complex deep learning models. Specifically, using TensorRT and cuDNN as opposed to a conventional CPU enabled us to detect, segment, and classify objects at a much faster rate. We were able to support this faster rate with the help of GPU, thus helping us unlock a more sophisticated on-vehicle data compression system, offloading CPU work to the GPU while also reducing I/O pressure on the vehicle.

Though L4+ self-driving technology is still very much on the road to full maturation, we are excited and encouraged by the progress we have made in China. We welcome these complex, unexpected conditions because they push us to continually challenge and optimize what we have built. We look forward to bringing this technology to China, the US, and beyond.