AI has made systems of a self-driving vehicle much smarter with less human cost.
At the Auto China show in Beijing, many of the new car models on display are equipped with autonomous driving features.
Plenty of top Chinese AI companies have developed autonomous driving software for cars. One such company, SenseTime, says the industry has hit a challenging bottleneck.
In order to let the autonomous driving system correctly recognize and analyze what's been collected, and execute corresponding driving maneuvers, developers need to write lines and lines of code to set rules for each possible scenario.
This is attainable for less complicated locations like highways, but on city roads where human actions are much more erratic, it's hard for programmers to keep up.
"There are many complex scenarios which cannot be defined by rules. Meanwhile, the design for every scenario requires thousands of program engineers and a very long time," said Wang Xiaogang, co-founder and chief scientist of SenseTime.
So, at the auto show, SenseTime presented a different solution: instead of improving various devices like LiDAR sensors and radars, its system relies mainly on video cameras.
The work of on-road rule-setting is relegated to an AI system, which condenses all visual input into something easily understood by an algorithm.
"In our self-driving systems, each module, from perception, fusion, to positioning and control, has been integrated into a single neural network. This has significantly reduced the workload in code development," Wang said.
The challenges of this approach are obvious too. Like many AI systems, a neural network needs to be constantly trained to self-improve, and that requires data.
"In the future, data is absolutely key. You need to have a lot amount of data to just train your model," said Cui Dong, managing director of Boston Consulting Group.
Experts say going forward, the global smart car race is likely to go hand in hand with the race for AI prowess.