Firmware

Tesla Autopilot ultrasonic sensors handle difficult merging event

Posted on

This was an unusual test of the Autopilot (AP) that was completely unintentional on my part.  My goal for the day was to video the lane changing feature of the Autopilot, and that’s why my cameras were running.  Even so, this situation provides an interesting follow-up to my previous video on the ultrasonic sensors as experienced in the perpendicular parking feature.

The perpendicular parking test, which I repeated about eight times in the same parking spot as used in the video, convinced me that the sensor-software combination is very accurate and reliable.  If I hadn’t had that prior experience, I may not have left the AP to carry through uninterrupted in the present video.

The video begins with a red truck passing normally while the two lanes continue well ahead of his passing location.  Following that truck, the driver of a Toyota decides to pass, but his decision to do so is too late.  Well before he reaches the rear of my car, the lane-marking between the two lanes has disappeared, and his pass will take place entirely in a one lane section of road that is predictably narrowing to a standard lane width.

 

In a situation where two lanes merge to one, the normal behaviour of the AP is to hug the outside lane and, after the lane-divider marking is gone, allow itself be shepherded by the converging outer lane-marking toward the centre lane-marking as the road narrows.  I’ve noticed in the past that this situation can be complicated by other traffic.  Sometimes I have taken control.

In this case the AP was steering as expected.  The Toyota then presented it with a vehicle closing on the Tesla’s left side while the lane-marking was closing in on the right side.  The sensors were picking up both the Toyota and the lane-marking.

A similar scenario often occurs in slow moving traffic when the Tesla is in the centre lane with traffic on both sides of it.  If the lanes are narrow, the AP stays within the lane-markings as general guidance, but as the sensors pick up the vehicles on either side, it adjusts its position within the lane so it is more or less equal distant from each flanking vehicle.

And again, during the perpendicular parking test, the AP demonstrated a goal of finding an equal distance between the parked cars on either side.  In that case it even exited the spot for a new approach so as to achieve that goal.

In the present video, the AP strove to remain within the lane-marking while being very much aware of the closing vehicle on its left.  In my mind there is no question that the AP is programmed to give greater priority to avoiding contact with the vehicle than transgressing the lane-marking.

I suspect that if crossing the lane-marking to avoid the vehicle became imminent, the AP would have required the driver to take control, but this did not happen, and the situation resolved itself by the Toyota pulling ahead.

This performance of the AP shows highly sophisticated programming/decision-making.  If the AP were to defer to the driver in every case that could be seen as tending toward risk, then we as drivers would probably be called upon much too often to take over.  This video probably shows about the limit of the strategy of letting the AP continue up to the point of imminent danger.  At each point throughout, I felt certain that I had enough room on the right side, beyond the lane-marking, and also to the rear, that I could have comfortably exited the situation if called upon to do so.

This experience enhances my confidence in the capabilities of the AP software, the hardware it relies on, and the general strategy of allowing it to remain autonomous well into complex scenarios.

Tesla Autopilot Demonstrations

To Top