Mad Max and AutoPilot in 2019.5.16

I already mentioned about my first impressions of autopilot under 2019.5.16 were that AutoSteer became more tentative, cancelling AutoSteer much earlier than in earlier version; but between daily commutes on back roads, down I-90, and down to Brooklyn and back, I've gotten to test AutoPilot in a much broader range of situations than normal.

When I first tested AutoNavigation with the roll-out of v9, I remember being surprised at how suicidal its lane changes were. It both chose to switch lanes and would perform manoeuvrers, if you confirmed them, without regard for traffic in the adjoining lane. Since then, Tesla has been adding blind-spot detection (something that was fairly common when we got our family car from from 2012), providing visual alerts in the driver's dash display, and presumably taking this information into account with AutoNavigation (although it should have been being incorporated into AutoSteer already, at least since v9 was released, introducing AutoNavigation. AutoNavigation and AutoSteering on the highway in general are much improved since the initial v9 release. Perhaps it always did this under AutoNavigation, but the lane it will want to put you into (e.g. to continue your journey) is indicated in the driver's display. What I did not see happening with the initial release was slowing when necessary to merge with the adjoining lane.

I even enabled MadMax mode on our recent trip, after finding AutoNavigation ill-prepared for travel on the Merritt Parkway (CT-15). The initial MadMax mode I tried last year appeared to want to switch lanes for the most absurd of reasons, but this appears to have been toned down (or decision making improved significantly). The main behaviour difference when switching to MadMax now appeared not to be its desire to switch lanes, but its aggressiveness (the demo versions of AutoNavigation apparently listed this setting under "blind-spot threshold"). What I found MadMax doing now was effectively playing chicken with nearby lanes. It would begin manoeuvres, but swerve back into its lane if it didn't have room. If this succeeded in getting the trailing car to slow down, it would try again.

Despite this new found aggressiveness, MadMax did not fit-in with traffic on the Merritt; it simply wasn't aggressive enough in wanting to switch lanes, not to mention a desire to "get out of the passing lane" when it wasn't even reaching its set speed. Elon had joked about a proposed LA Freeway setting, which might have performed better (although to me, as an East coaster who's only travelled twice through LA, my perception of the freeways there are slow crawls). The AutoSteering system is not without other flaws, aside from AutoNavigation tuning, as well. When we crossed the Hutchinson River Parkway Bridge, a drawbridge which loses its marked lanes, AutoSteer became confused and rather than continue in a straight line, tried to steer into the lane to its right, even though an SUV occupied that lane (reminding me of an experience under 2018.21.9). I found the drawbridge experience odd, as the car has otherwise been doing much better with gaps in lane markings, able to make it through broader (and non-straight) intersections for which Massachusetts is notorious even where it was not able to in previous versions.

Perhaps the biggest improvement in AutoNavigation compared to the initial 2018.42.2 implementation, is how it handles exit ramps. The initial implementation did not slow down, and often began turning before the exit ramp requiring manual intervention to refrain from either hitting a guard rail or driving off the road. Under 2019.5.16, exiting actually seems to work, sometimes. The car seems to trust its standard vision/AutoSteer over what appeared to be a bias towards map-based pathfinding for the exit ramps in earlier iterations.