Google (Waymo) and Mobileye both are doing self-driving cars and they have different strategies,
|| L2 -> L3 -> L4
- Sensing surroundings
- Driving plan based on map and ML of good driving and the surrounding image of his car
- Sensing surrounding objects
- Driving plan based on ML on map, position/movement of the objects in the surrounding and mostly-used tracks on the roads
|Years of development
|| 8 years (from approx. 2009)
|| 18 years (from approx. 1999)
- 2,777,585 km driven in autonomous mode (till June 2016)
- Test based on pre-programmed route data
- Total 14 collisions: 1 program caused, others cause by human error
- Not be able to spot potholes or a police officer signaling the car to stop
- Mobileye has provide more 90% OEMs with its products
- The Tesla AutoPilot crash in 2016 was caused by Mobileye’s camera which, against a bright spring sky, failed to distinguish a large white 18-wheel truck and trailer crossing the highway
|Year to reach L4
Autonomous Level 4 is high driving automation at which a driver is still in the car and he can overtake the autonomous system.
Waymo also detects objects but focuses more on its deep learning to mimic a good driver. However, a good driver can handle “new” and not-experienced situations safely, which is the rare cases in ML database. Without enough such rare cases, ML will never learn well. To collect such cases, it faces two problems,
- ML resource will exponentially increase
- It will never achieve zero-fatality rate with limited cases for ML because there are always cases AI driver does not see/learn
On the other hand, Mobileye sets the goal of zero-fatality rate for its system. The fatality rate for 100 million Km (in USA in 2015) is 0.7 . It is questionable if it is a realistic goal to have zero-fatality rate for autonomous drive.
Furthermore, Mobileye claims that its approach is different from Waymo’s. I think they have big overlaps.