They've been "coming along nicely" for something like 10 years. Google's not really doing anything that hasn't been done before, but they are taking advantage of the lull in real developments in this field to exploit public forgetfulness.
We could make cars that drive themselves perfectly right now. We could have done it as soon as 5 years ago, maybe sooner. It is a simple matter of drilling special beacons into the asphalt every 6 inches or so and letting the cars follow the beacons. I witnessed robots driving themselves along paths like this in factories in person in 2003. What about at speed? How do they know about other cars? For that, every car needs a nav beacon that can be detected by other cars for at least a few hundred yards. When every car knows where every other car is, then each car can maintain minimum safe distance and when one brakes hard 5 miles away your car can take measures to compensate. Are you ready to live in a world where <government entity>/<kid with radio shack hardware> can see where your car is at any time of day? Most people aren't. Regardless, the cost of this project would be in the tens of billions of dollars for just the interstates; doing it along city roads would increase the cost further. Not to mention all road maintenance would have to involve drilling special maintenance beacons into the road first, and typical road damage like potholes would suddenly become far more life-threatening. And the whole system is at risk of non-car obstructions; it's going to require a pretty slick LIDAR or other rangefinder to catch cattle in the road.
The cheap technique is to strap a camera onto the car and have it follow the road. Computers can do tens of billions of calculations per second but it's all done serially; even the great supercomputers are only capable of perhaps a few thousand parallel simultaneous calculations. On the other hand, your brain's processing several millions of pixel data in parallel at ~24 frames per second while also analyzing audio and manually adjusting for the lag between visual processing and audio processing. The brain has very specialized circuits for pattern recognition in image data that we have yet to understand, let alone replicate. It's also capable of generating ideas; something computers find tricky. It turns out this task that's really easy for people is incredibly difficult for computers.
It only works if the lines are very high contrast against the roads. Usually that means very black asphalt and very white paint, but if the paint's reflective enough it'll work against lighter asphalt. In urban areas, curbs and parked cars might serve the purpose of lines. This is a problem. On many roads, the paint has low enough contrast that humans have difficulties finding the lines. Sometimes the road crew screws up and the lines are curved or shaky. Many rural roads have no lines at all. Wear and tear over time darkens and damages the paint. Road maintenance often diverts you to shoulders or places where there are no lines or the lines are on the wrong side. The camera might have a hard time recognizing other cars; again the computer tends to rely on highly reflective surfaces like license plates. What about new cars with no plates? What about people that lose their bumpers and put the plate in the window? What about people with no plates? Traffic lights are another problem. Sometimes an accident causes police to direct traffic at a light; how does the computer recognize this? Again, we find ourselves needing a system of beacons to assist the computer, and the expense is intractable. Again, a network of the location of every car on the road would be most helpful, and this is a deep privacy and safety concern.
GPS doesn't help. It's not accurate enough to keep you in a lane, and I haven't used one in Austin that wasn't wrong about where to turn into businesses by a few hundred feet. I don't trust that to stop me from running through a red light or into another vehicle.
No, I'm afraid the only cheap and feasible answer to our transportation problem is to give up the "American dream" of living in remote suburbs (this was a system pitched by GM to sell more cars!) and live in more dense areas served by public transportation. Cars that drive themselves are at least 25 years away in my vision, and every year that goes by without commodity chips (think < $1 for the equivalent of an early Pentium 4) pushes that out further.
I don't want to be a downer, but I also hate seeing Google steal thunder for doing research that's already been done and encountering challenges that have already been well studied. Autonomous cars aren't going to be exciting until they drive themselves down Route 66 and do the street view for towns like Soso, MS or Ellisville, MS, where the roads are poorly maintained and don't have lines at all.
On the flip side, there is some value to a juggernaut like Google getting interested in it. If certain standard road designs are followed, it becomes much easier for cars to drive themselves. If Google could toss some money to lobby for very specific changes to how new roads are built all the way down to the city level, they could make it much easier to solve this problem. I'm not excited that they're doing the research as much as I'm excited that they think it's important. They can grease a lot of gears.