They're monitoring human drivers to see how they interact with the world and then fine tuning based upon failure states. So it starts broad and then they narrow the focus to sculpt the system into operating appropriately. So in that video they discuss the fact that 99.5% of drivers don't come to a full a complete stop at stop signs (rolling stops) so they query the fleet and show it examples of the .5% of scenarios where the vehicles do stop at stop signs in order to get the system to recognize that as the appropriate behavior.
They have 2.5 millionish vehicles on the road in their primary development area (North America) and can pull up scenarios like a vehicle pulling a wobbly trailer, or a ladder flying off the back of a truck and show the system how to respond to those scenarios through actual video and through simulation of those events. They also have ~500,000 beta testers that accumulate around 1,000,000 miles/day and they get fed instances of disengagements from the fleet so they understand where the system is not performing as it should.
From a regulation perspective they operate at SAE Level 2, which allows them to operate on that large of a scale, because the liability remains with the driver, rather than the system. In practical terms Autopilot is never designed to be any more than a SAE level 2 system, because it is restricted to highways, but FSD Beta (different than autopilot) is more like a SAE level 3 system, but operates as a driver assist system. As it improves they can extend the duration between interventions of the system until it gets to the point that there's an intervention every 10 or 100,000 miles (whatever regulators expect) between interventions and they can then deploy a level 4 or 5 system without ever officially going through the level 3 stage.
It has been a pattern of two steps forward, one step back for awhile, so they're continually improving the system it's just a matter of how long it takes to go from 5 miles between interventions to 10, to 100, to 100,000 (actual time between interventions is unknown, but you can watch videos of people going long periods without intervening on YouTube).
Tesla has 3 different driver assistance tiers.
Autopilot - comes with every vehicle, basically just adaptive cruise control and lane keeping
Enhanced Autopilot ($6000): adds lane changes (I.e. it will navigate around slow moving traffic), and navigation so it will take exits to follow your plotted course on highways.
FSD Beta ($12000): Adds driving on city streets. So you can get in your car, back out of your driveway, and activate the system and it will stop at stop signs, stop lights, make left and right hand turns and right on red/left across traffic etc etc. The driver still maintains liability for the actions of the vehicle in the same way a person using lane keep/adaptive cruise control is responsible for the vehicle.
Tesla is adding about a million vehicles to their fleet right now, with a 30% growth rate every year to that number (previously 50%, but they've mostly saturated current models and won't be adding another mass adoption vehicle until 2025/2026ish). So they'll go from 2.5 million vehicles collecting data, to 3.5, to 4.8 million etc being able to add more and more data/edge cases to train on.
.
