Definitely new for me! Had my first commute on 40.2.1 today and as I suspected, but hoped would not be, did indeed happen. While using NOA in my 38 mile stretch of being in the HOV/PeachPass lane today the new “feature” that slows your car when traveling at a substantially higher speed totally screws up using NOA usage. Several times I was traveling at ~60Mph and the shaded arrows showed up in the adjacent lane and slowed my speed to <45mph. Traffic in front of me left me and the traffic behind me immediately went into rage mode. I had to take over and press the accelerator both times and after the second instance just clicked “off” NOA for the duration.
Is the HOV/PeachPass lane separated with solid lines or just regular dashed lines? If solid with occasional breaks to dashed, does the "Adjacent Lane Speed Adjustment" kick in only when it temporarily switches to dashed? In any case, the highways near Tesla HQ in Palo Alto have HOV lanes only separated by dashed lines, so I'm sure the Autopilot engineers there will run into this undesired slowing down themselves (and hopefully adjust the behavior): Google Maps
It is both, changing to dash lines where entry exit is allowable. The decreased speed actually happened when in the Solid Lane Double Lines portion tho. Traffic usually getd closer in speed at the dash point anyway which is probably why it didn't seem to happen there:
Elon thinks their AI is much smarter than it is in practice. Unimproved experience with v10: - car likes to "flinch" when there is a car in the adjacent lane a little too close to the line - auto wipers are worse, I had to intervene manually numerous times yesterday - slowing down when in the faster lane defeats the whole purpose of an HOV lane - NOA will try to move into the right lane when there is a gap, despite that lane being full of cars 100-200 yards ahead. This is the sort of obnoxious driving style I find really annoying when I'm in the left lane. - Cruise control is much more jerky than in my 2016 S90D. My wife can immediately tell who is driving. - Other than knowing that an exit is required 1-2mi ahead, NOA makes poor decisions. - Smart Summon is amusing, but it's much faster, and healthier, to walk to your car than wait for it to go through its very slow dance to drive to you. I have no use for the arcade games and videos but I'd really like to have the smooth cruise control and AP that was in my 2016. I think Tesla has very little to show for four years of "self driving" software development. Elon loves to sell the sizzle and not the steak, but there needs to be some steak under that sizzle.
Piggy backing on this comment here. I don’t remember what update it first appeared in, but I god damned HATE that the car is afraid of anything it thinks is a large vehicle, which is like 90% of the vehicles on the road. Truck in the lane to the left of me? Better swerve to the right of my lane!! Little van to the right of me? Yup, that’s definitely a danger worthy of cowering to the left of my lane. These vehicles are driving normally too, so my car is just running away from them for the sake of running away. I bet the cars around me appreciate the fact that I’m the one swerving around like a madman. It’s plain stupid and unsafe. At least when “truck lust” was a thing, it was just for a single type of vehicle. Now we’ve got “vehicle cowardice” for most of the cars on the road, further reinforcing my opinion that autopilot is the equivalent of a timid teenager on the freeway for the first time.
Absolutely. Raven S hw3. 40.2.1 Accelerates up to max speed even approaching junction / roundabout (in UK). With wet road surface (after rain) but otherwise perfect conditions, reflections off road create random traffic cone detection (mis-detecting brake light reflection). Parked cars at roadside often seem to be either ignored (vehicle makes no attempt to avoid) or treated as queue of stationary traffic. Off-ramps or anywhere with non-uniform road marking especially cross-hatching (what's that called outside of UK?) treated almost randomly. Once it nearly pulled the car back onto the main highway when car was already on the off-ramp. Cyclists show fine on display, but no attempt to pass. Just slows behind cyclist. (I haven't been able to fully explore this one, but it felt like quite unpredictable behaviour, especially if it sees a cyclist on the sidewalk. Various heavy fantom braking followed by fast acceleration. Often seems to drive far too close to curb leaving no chance / room for driver intervention if needed. Uneven breaking / juddering of brakes when applied by AP while regen is limited during warm-up. Suggestion of incorrect response to cars parked in lay-bys (don't think they exits in US?) immediately adjacent to main carriageway. Seemed to 'almost' trigger emergency breaking once or twice. Parked vehicle showed as red object on IC display. Brakes applied hard for split second, message about 'take control of vehicle immediately' . The reflection issue and response to cyclists on the sidewalk were of especial concern. If the car is to respond at all to traffic cones, what will it do if it 'sees' reflections off the road surface that it detects as cones?
Seems like a lot of your listed concerns are from using Autopilot on "city streets" and the similarly poor behavior happens in the US as well. I guess the main difference is that there's a higher percentage of "city streets" vs (divided and access controlled) "highways" in the UK vs US. Just making sure, Navigate on Autopilot is not active for you and/or you can't set the speed higher than 5mph over the speed limit?
Yes, far more smaller streets. When it's been raining, the reflective road surface causes havok for AP. I lived in San Diego for 15 years so I'm familiar with SoCal driving! No NOAP. (It is a FSD car though ). I set speed typically 5mph houghoughh below posted limit because that makes it less scary! It is the jumpy, twitchy unpredictable nature that is actually not only frustrating but genuinely dangerous. Street layouts are generally more compact here, and most probably far more likely for objects to be seen by AP as significant when in fact they are on the sidewalk. (got to be careful with terminology as in UK 'the pavement' is where pedestrians walk ie: sidewalk!) When I drive manually I feel like my decision horizon is typica whatlly between 50 and 200 feet ahead, or longer on freeways. Tesla AP seems to set its horizon about 20 to 30 feet ahead and is quite happy changing its mind completely about what it thinks it 'sees' without regard for saw milliseconds earlier!
That's a great analogy. I'm not certain this is a bad thing though. Tela is literally training an AI to drive. It's got to go through steps to gradually learn.
On Sunday, I used NOA for the first time since installing 40.2.1. The lane changes were much smoother and quicker than when I was on pre-40. software. Overall, a very noticeable and welcomed improvement.
It's only a bad thing because Tesla has been telling us they are learning for four years! A driver who starts at 16 makes a lot more improvement by the time they are 20 than Tesla has in the last four years.
I would have more confidence in this thing if they didn't have the 3D ADAS visualization. It's really cool when it works and has come a long way but you still see adjacent cars phase out of existence, or the depth estimation can't decide what lane they're in. Long vehicles that cross multiple cameras just go totally bananas on there. They enabled visualization of oncoming cars in a recent update and it shows maybe 33% of the cars that pass by.