This field of research is far more advanced than I thought, actually.
For instance, in the minute 25:12, they show a graph that compares how often the time to collision among a human driver (a professional one, actually) and the self-driving car. For small reaction times, the self-driving car outperforms the humans by far. It is of course expected, but it shows that driverless cars are not that distant in the future already.
The biggest issue is how to avoid those outlier situations, such as a plastic bag being interpreted as an obstacle and making a mess. Fixing these is a huge challenge.
A plastic bag could also fly right onto the LIDAR, completely obstruct it, and it wouldn't matter what the software is capable of. These cars are going to need some sensor redundancy.
The Velodyne LIDAR Google uses costs somewhere near $75,000. I realize these sensors are new tech and are likely to nosedive in price, but having two is currently tremendously expensive.
There needs to be a LIDAR unit that performs as well for a lot less before these are going to ever hit regular people.
For instance, in the minute 25:12, they show a graph that compares how often the time to collision among a human driver (a professional one, actually) and the self-driving car. For small reaction times, the self-driving car outperforms the humans by far. It is of course expected, but it shows that driverless cars are not that distant in the future already.
The biggest issue is how to avoid those outlier situations, such as a plastic bag being interpreted as an obstacle and making a mess. Fixing these is a huge challenge.