The short answer: Pretty well, but don't rely on it.
One of the biggest concerns motorcyclists have about autonomous vehicles is how well they will detect us and avoid us on the road. Crashes involving autonomous cars have already happened, and the American Motorcyclist Association has taken a strong stand on the issue. A recent video by Scott Kubo shows us what it's like from the car's perspective as he demonstrates how well his Tesla Model 3 detects motorcycles, and in some cases fails to detect them.
This video takes place in California, the one state in the union where lane splitting is legal. It's a little bit of an edge case, but it makes for a great demonstration of how well the Tesla senses motorcycles in one of the most difficult situations it will encounter. You get a great view not only through the windshield but also in the rear view mirrors as well as the Model 3's rear camera display. The sound is also in stereo, so put on some headphones to hear the bikes pass on either side.
It's tough to see in the unedited footage, but later on, there is a close-up of the Model 3's center display that shows us what the car sees. Autopilot differentiates between cars and motorcycles, as can be seen by the use of different icons on the display. Indeed, sometimes Autopilot detects the motorcycle passing directly in between lanes, exactly as it should.
Autopilot isn't perfect, though. Sometimes it thinks the lane splitting motorcycle is passing directly through the car in the next lane. This isn't a big deal, because at least it's detecting a vehicle there and won't crash into it. Autopilot will not move over within a lane to make more room for a motorcycle to pass, but riders can deal with that. Kubo also shows himself deactivating Autopilot briefly to move over and help bikes get around him.
From Kubo's observations, it seems that Autopilot isn't necessarily good at detecting motorcycles with its cameras, but the car's ultrasonic sensors still do a good job of it. Perhaps it's just as hard for computers to see motorcycles as it is for human drivers. If a motorcycle approaches too rapidly, even that data can't be processed quickly enough to detect the bike properly before it's already gone. Kubo mentions that new computer processors due to come out next year will be ten times faster than the current ones, which may help.
Another interesting point that Kubo makes is that he often first notices motorcycles on the road by their exhaust notes. No self-driving system currently takes sound into account. This would likely be rather difficult to do since there are so many variations on engines and exhausts for motorcycles to program into a computer's memory banks, but it's a neat idea to possibly develop in the future.
The points to take away from the motorcyclist's perspective are simple. Handle autonomous, or semi-autonomous cars like Teslas, the same way you would handle a human driver. Don't trust the car to automatically detect you at all times, just like you shouldn't trust any human the same way. Most importantly, especially in a lane splitting situation, don't ride significantly faster than the traffic you are passing. Not only will Autopilot be unable to detect you, neither will human drivers.