Third Day: Autopilot

With three days of driving, I am starting to get a better feel for how Tesla's Autopilot behaves. When I first started using Autopilot, there was an initial hesitance which melted into overconfidence. After the third time you find the car drive into on-coming traffic, or the fourth time you find the car accelerating into a tree as it disengages aproaching the apex of a turn; you start to get a feel for its limitations. Admittedly, Tesla's directions strongly suggest only using Autopilot on essentially the Interstate system, or similarly controlled roads; but unlike General Motor's Super Cruise, it does not actually restrict you to such a narrow network.

Full Self Driving Hardware

My Model X, Pensive, had the Full Self Driving package on top of the Enhanced Autopilot. This means, technically speaking: that rather than the four cameras provided with Enhanced Autopilot, the car has eight (Autopilot 1, or unhenhanced Autopilot, had one camera). Functionaly and effectively, it means that I paid Tesla $3,000 to make my car more complicated since no released software makes use of the extra four cameras. The four cameras in use allow the car to see 820ft straight ahead within a narrow cone of view (and 195ft with a wider cone of view), as well as 165ft rearward. Unlike other companies (and research institutions) looking to develop self-driving cars by leveraging lidar, Tesla (synecdochically via Elon) insists that cameras and SONAR are all that is necessary to achieve self-driving. Lidar provides much higher resolution than SONAR, and more accurate measurements of velocity than video cameras; but it does have drawbacks, namely environments where light would be scattered (e.g. rain, snow, fog, dust) effectively blind the technology. This is one reason why Elon refers to lidar as a "crutch"; even with a principa lidar system, there needs to be other technology to handle heavy rain. (Heavy rain at night, where neither lidar nor video would perform well, would presumably mean relying on bat-like echo-location, i.e. SONAR?)

Autopilot User Experience

Tesla's Autopilot system encompasses a lot of different, but related, technologies. The most interesting of which is usually its Autosteer system, which is its lane-keeping system. Unlike a lot of other cars which have a lane-keeping assist system or lane-departure warning that vibrates or otherwise alerts when the car appears to be crossing a dividing line without signalling, the Tesla Autosteer system included with Enhanced Autopilot is intended to keep the car (roughly) centered in the lane without input from the driver: any significant input from the driver will deactive the Autosteer system. This is an interesting aspect (although similar to other semi-autonomous systems, like Volvo's Pilot Assist) since the car and Tesla continue to rightfully assert that this is a driving aid and not fully self-driving, so the system attempts to verify that you've actually got your hands on the wheel.

It's not immediately obvious from the literature that Tesla provides how the system does this; such banality isn't something that most people look for when reading the virtual glossy pages of a brochure-cum-Website. While the Super Cruise system found in Cadillacs uses eye-tracking, and the Model 3 has an inward-facing camera, Tesla relies only on a torque sensor in the steering wheel (as does Volvo's implementation). The assumption is that as Autosteer drives the car, the driver will naturally provide light resistance which is registered by the sensor, thus indicating that the driver has at least one hand on the wheel, and is therefore paying attention and alert. Frankly, this is wrong, as anyboy using these systems has probably experienced. If you're driving on flat, straight, roadway, there will be very little adjustments the wheel makes so very little to record counterpressure from the driver. If you naturally have a light grip on the wheel, treating the wheel the same as if you had released it straightening out of a turn, it won't record your hands being there. If it warns you enough times, it will eventualy disable itself. To keep Autosteer happy, you might, like me, develop a nervous tic of rocking the wheel periodically. (I find the sport steering setting best for this, although standard isn't too bad either; but frankly I can't tell whether it has an impact compared to the twitchier comfort on whether Autosteer recognizes the movement as significant enough to disable.)

An obvious place where Autopilot helps is when you're in stop-and-go traffic. In fact, there are two places where most people will use Autopilot for the first time: stop-and-go traffic, or when there's nobody around. While stop-and-go traffic lets you immediately get acclimated to the technology at low speed (you'll start when you're essentially stopped, but to enable it you'll have needed to lift off the brake which is a little annoying), you'll start to understand its limitations when you're testing it on wide open roadways. It's when the road is wide open that it will be going at the full speed you set, so you can see how it handles certain types of roads. My favourite is the winding back-road where Autosteer is limited to 35 mph, but non-Autosteer cruise-control is still set to, say 45 mph. As you approach your turn, you think to yourself huh, I would have started turning by now just before the car disengages Autosteer and accelerates up to the cruise-control speed of 45 mph and right towards a tree.

An interesting tidbit is that while Autosteer claims to be limited to certain speeds, such as 35 mph in the above example (generally, the local speed limit +5 mph), you can modulate speed via the accelerator pedal without disabling Autosteer. This will, for instance, allow you to go faster where it is safe, and let-up as you approach a corner to let regenerative breaking bring the car back to a speed where Autosteer is comfortable taking the turn.

Identifying Problem Situations

You quickly identify some of the problem situations which confuse Autopilot. You get a feel for which corners are going to be too sharp (and wonder why the car doesn't). You start to identify which intersections the car will get lost in, and drive into the on-coming lane (although this seems to be harder to generalize than the somewhat sharp corner conditions).

Essentially, the rule to driving a Tesla on Autopilot (and driving generally) is knowing your limitations. This means listening to what the system tells you, but also what you're seeing with your own eyes. The car naturally does not take corners at an optimal angle, preferring a late entrance, and treats the car as much less agile than it is (at least in good conditions). If you look for areas that are unmarked that might present a problem, narrow or rapidly changing areas, or curves which look to be sharper than for which the car is accustomed; Autopilot is manageable. In some of these cases, such as sharper curves, slowing the car down via the cruise-control-down toggle may suffice; but practically, it's easiest to just re-enable on the other side of the curve.

Non-Autopilot Issue: Sound

Completely unrelated to auto-pilot, today I noticed an issue with the sound system. There was a definite, intermittent, change in sound quality and I spent most of the trip into work trying to isolate which speaker(s) were causing it. Somewhere along the trip home, I thought I'd isolated it to the front-left speaker, seemingly going in-and-out; and while it seemed like the sort of behaviour that would be caused by a loose connection, the timing seemed suspect, as if it could be a software issue instead.