It's a lot cheaper than it used to be. When we did our DARPA Grand Challenge vehicle ten years ago, we had about 3 degrees of heading noise, which kept the system which integrated LIDAR data into a terrain map from reliably mapping. The system had a MEMS gyro with too high a drift rate, and a magnetometer compass for long-term correction. But there were enough magnetic fields from the onboard equipment to mess up the magnetic compass. The software could detect trouble, and would stop, tilt the laser head to scan the terrain ahead and get a clean image, then advance. So the vehicle kept stopping to rescan.
We looked into better heading gyros, but they cost about $20K at the time.
Applications for IMUs include dead reckoning (working out your position based on rotation and acceleration measurements), orientation sensing (figuring out which way is down is much easier than dead reckoning, because you can just sense gravity), and compensating sensor readings for robot motion.
Some cheap sensors aren't very precise; you won't notice if your phone thinks it's rotating two degrees a minute around the vertical axis, but a robot will if it moves for five minutes then it's ten degrees off course. I haven't looked at the market for a long time - I suspect the cheaper IMUs are better nowerdays. Eight years ago they didn't offer very good precision and the open source software could be very fiddly to get working, if you could get it working at all.
The OXTS RT3000 has impressive specs - 1cm accuracy in world coordinates. That's better than you need for most robotics applications that aren't running blind. The Trimble AP10 [2] has the same functionality at much lower cost, but is not as accurate.
A lot of progress in robotics comes from component improvement like this. There was a time when just getting a wireless data link to your robot was a big deal.
[1] http://www.analog.com/en/products/mems/mems-gyroscopes/adxrs... [2] http://www.trimble.com/gnss-inertial/AP10.aspx?dtID