Video from Taiwan reveals a disturbing Tesla
The video shows the event from several angles, and raises several questions:
- Why does Tesla autopilot not perceive such a large obstacle on the road as this, with its use of cameras and radar?
- Why does the Tesla emergency braking system not brake for the driver of the truck, who is standing in the lane in a misguided effort to direct cars away from hitting his truck? When the Tesla does not stop, he jumps into the shoulder and is unharmed.
- Would maps or LIDAR have prevented this accident?
- How much attention did the Tesla driver pay to the road, and what bearing does that have?
- Was Autopilot actually on as the Tesla driver claims?
We can begin with #4 — Tesla’s Autopilot is a driver assist system and it requires drivers pay attention to the road at all times, and puts responsibility for accidents upon those drivers. Drivers must wiggle the wheel from time to time to prove they are keeping hands on it, but their gaze is not tracked as it is in some cars. This driver could not have been paying much attention — the truck is as plain as an obstacle can be on a fairly empty road — though he does hit the brakes just before impact. So there is clearly driver fault here. However, Tesla regularly says that their systems are very close to being capable of real full-self driving, and a system which does this is nowhere close to being ready for that.
This failure to perceive is so glaring that one has to wonder if the Autopilot was actually on. Tesla declined to comment on that. However, given that the driver clearly was not attentive and the car drove fairly straight in the lane prior to the crash, it seems it is likely to have been on. In addition, Tesla’s “Automatic Emergency Braking” is generally never off unless manually disabled, as are the collision warning systems.
Missing a Truck
The answer to missing a truck is reasonably well known. Computer vision systems recognize things they have been trained on. Seeing the roof of a truck on the road is not a common event. Tesla’s image classifiers probably have not trained extensively on trucks lying sideways on the road. The back of a truck they will identify, and by now, perhaps the side of an upright truck.
The second issue is radar. Generally, the radar in the Tesla will have received strong reflections off this truck. (Note the roof is not metal, though.) However, these radar returns would indicate the truck is a stationary object, just like the hedge in the middle of the road, which also will be giving radar returns. Radar resolution can’t always tell something in the median of the road from something in the left lane. In addition, the road curves to the right here, and so there are median objects directly along the path of the lane, and the car must not brake when detecting those.
Below, we’ll explore how maps could resolve that radar question.
Missing a pedestrian
One reason to wonder if Autopilot was on is that the vision system is well trained on pedestrians and should now be at the point of almost never missing them. Pedestrians are not expected on freeways (which is why human drivers often hit them) but they should still be detected. And there also would have been radar from the pedestrian, but he was not moving up or down on the road and the same radar problem emerges.
Even Tesla’s AEB system (which is usually never off) should have reacted to that pedestrian. If it did, it should have alerted the driver, and hit the brakes itself much sooner.
Maps and LIDAR
Sadly, once again, two technologies Tesla has deprecated as “crutches” — LIDAR and detailed lane-level maps, could have saved the day. LIDAR would have detected the truck very clearly and triggered braking quickly, there is little doubt of that. Even the most basic non-scanning LIDAR would have done that.
Maps would have possibly spoken of the radar profile of that section of road. While they would have told the vehicle to expect radar returns from the median, it would know their character. The radar returns from a big truck like this would be stronger. In particular, the fact that the truck extends over 3 lanes should have meant enough to detect that something was stopped in the lane to the right, a sign for caution. Maps would have revealed the curve in the road ahead which puts radar targets in the median directly ahead, but they would also report the exact distance to these targets which would be much further than the distance to the truck or the truck driver.
It is baffling why a large flat surface and the large body of metal did not provide sufficient radar signature to trigger Tesla’s radar system.
Driver Attention
Tesla’s instructions require drivers to pay attention to the road. Some don’t, however. Some products attempt to prevent that by monitoring the gaze of the driver, and warning one looks away from the road for too long. Tesla has decided to not use that approach. It points to a good safety record for Autopilot as evidence this is not necessary. (Tesla refuses, however, to clarify what their safety statistics actually mean, because they do not say what an accident is, or what the record is on different road types, in their comparisons.)
The debate about how much driver monitoring is needed will continue for some time.
As a driver assist system, Tesla Autopilot is not expected to catch everything on the road, and so if drivers do not pay attention, there will be accidents. But missing a giant truck and a pedestrian is a bit much for a system which is also the foundation of a purported “full self driving” system. This reduces confidence in when such a system can actually ship. It is unknown if this Tesla had Tesla’s new hardware able to run the more advanced autopilots. Tesla declined to comment.
