Not from scratch! We use lots and lots of open source (<3 ROS), in addition to secret sauce we've developed over the past year.

As long as human-driven vehicles are a significant fraction of the vehicles surrounding an autonomous vehicle's miles driven metric, I'm not sure a technically-driven approach to edge cases is gating mass adoption. Humans likely don't have a very good track record with those edge cases, but the ones who deal with the aftermath of all those edge cases on an actionable basis are the legal and insurance industries. Luckily, they have copious documentation of those edge cases; unluckily, that documentation is not in a readily-utilizable format for computers.

Autonomous software development might finish 90% of the technical challenge, only to discover 90% remains to encode legal and insurance precepts before mass adoption is allowed because the safety outcomes bar to clear for autonomous driving will be higher than manual driving. Once you reach edge cases where the only physically-available options are all bad outcomes (where either riders get hurt, or people outside the riders get hurt), until sensor and simulation technology get good enough to evaluate the degree of how much each party gets hurt to discover the minimal injury option, we might be forced to make do with developing software that chooses an option that presents the least risk exposure to legal and insurance liabilities according to the governing jurisdiction. If so, then that would be a real mess to wrangle.

Hopefully, legal, insurance and regulatory frameworks around the world will recognize in autonomous vehicles with LIDAR+radar+sonar+vision+sound sensors, with the kind of testing the big players are performing, mortality per million miles driven are bound to be lower than human drivers, and not engage in a perfect-enemy-of-good position.

If politically-influential stakeholders put up barriers to entry in the form of demanding better-than-human perfection in outcomes, then we could wait a long time as players switch to alternative go-to-market plans. One possibility might be embed the sensor and software technology into manually-controlled cars as accident mitigation features, but simultaneously collect massive amounts of data to refine the edge case handling characteristics, and gradually ease into self-driving.