And just in case other people come here later and don't follow Kim's links and read through that thread, his comments were:

We've been running various experiments with mcomb to find out how the biggest problems could be overcome.

Initially, we started with the Tiger data (that is freely available for the whole US area). The problems with that data is that it's fairly inaccurate and it's missing one-way info/turning restrictions.

Recently, I've looked into the USGS DLG data (also freely available for the whole US area) which has better accuracy and has one-way info. But since the data lacks street names and numbers, it is essential to combine these two maps together. However, the map conflating seems to be quite problematic. I can get about 80% of the streets mapped correctly, but the rest gets mapped to incorrect streets.

This is a common problem in map conflating and without a relatively complex algorithms and manual tuning afterwards, it is hard to get good results.

Read more here and here.

Another problem with the one-way that can be extracted from the USGS data is that I haven't been able to figure out which way a road is one-way. If a road is marked as one-way, it's not necessarily always one-way towards the same direction as the data is digitized.