Thread by @ceptontech

1. THREAD: Advanced driver assistance systems (#ADAS) based on cameras + radars ("Level 2" or L2 ADAS) are not and won’t ever be “full self-driving”. Using examples in this thread, we’ll illustrate why.

2. This thread was prompted in part by a recent @jalopnik article on Tesla users’ concerns over viability of “FSD” (“full self-driving”)

jalopnik.com/tesla-owners-t…

3. It’s also partly prompted by an article in the @nytimes by @nealboudette that discussed ongoing Federal investigations into accidents involving Tesla vehicles

https://twitter.com/dealbook/status/1374710998221922306

4. Experts have weighed in before to point out that fully autonomous (aka “Level 5 or L5”) vehicles – or for that matter L4 vehicles – are not feasible using just cameras and radar; for instance see: @PeterCBigelow @samabuelsamid

https://twitter.com/PeterCBigelow/status/1286040277539528706

5. It is obvious that L2 ADAS absolutely helps improve safety, reduce accidents and save lives - and more adoption would be welcome – see @consumerreports

consumerreports.org/car-safety/cad…

6. However, a system that improves safety doesn’t automatically become a self-driving system and it may not be obvious to many people as to *why* L2 ADAS is not able to support fully autonomous driving. That’s what we’ll try to illustrate here.

7. Let’s start with an important fact: L2 ADAS systems in modern vehicles require full driver attention to take over when automation fails (this also applies to Tesla Autopilot and “FSD”) @strngwys

caranddriver.com/news/a35785277…

8. When might an L2 ADAS system in a car run into an issue? Anytime it is subject to known limitations of camera and radar technologies. Here’s an incomplete list of examples.

9. Pedestrian Detection, Ex. 1: Difficulty detecting a pedestrian on the road in front of the vehicle – AAA found many vehicles unable to do this well, sometimes even in daylight

10. Pedestrian Detection, Ex. 2: Difficulty detecting a pedestrian on the roadside (~08:52 mark, Tesla)

11. Pedestrian Detection, Ex. 3: Difficulty detecting a pedestrian on the road (Tesla)

12. Bike Detection, Ex. 1: System confused by bike / bike disappears briefly (~07:22 mark, Tesla)

13. Bike Detection, Ex. 2: Unable to detect a bicyclist until very close and can’t easily navigate around the biker (~08:39 mark, Tesla)

14. Lane Detection, Ex. 1: Difficulty keeping the vehicle in lane with older roads and bright sun (~03:28 mark, Tesla)

15. Lane Detection, Ex. 2: Difficulty with making a lane change under direct sun (~00:30 mark, Tesla)

16. Lane Detection, Ex. 3: IIHS found several vehicles had difficulty performing lane keeping in curved and hilly roads iihs.org/news/detail/ii…

17. Lane Detection, Ex. 4: Difficulty staying in lane in hilly and curved roads (~01:05 mark, Tesla)

18. Lane Detection, Ex. 5: Difficulty staying in lane in curved roads (~02:57 mark, Tesla)

19. Lane Detection, Ex. 6: Difficulty staying in lane in curved roads (~04:19 mark, Tesla)

20. Lane Detection, Ex. 7: Difficulty staying in lane in snowy roads (Tesla)

https://twitter.com/binarybits/status/1354917236419551237

21. Vehicle Detection (Curved Roads), Ex. 1: Delayed / incomplete detection of vehicles around curves (Tesla)

22. Curb / Road Edge Detection, Ex. 1: Detecting curbs/road edges can be particularly difficult (~07:47 mark, Tesla)

23. Curb / Road Edge Detection, Ex. 2: Difficulty detecting a raised curb (~00:34 mark, Tesla)

24. Curb / Road Edge Detection, Ex. 3: Difficulty detecting a raised curb (~03:45 mark, Tesla)

25. Curb / Road Edge Detection, Ex. 4: Difficulty detecting the edge of the road (~01:52 mark, Tesla)

26. Free Space Detection, Ex. 1: Braking hard when approaching/passing vehicles on the side of the road despite sufficient driving space (~06:58 mark, Tesla)

27. Free Space Detection, Ex. 2: Braking hard when approaching/passing vehicles on the side of the road despite sufficient driving space (~05:39 mark, Tesla)

28. Free Space Detection, Ex. 3: Braking hard ahead of oncoming vehicles that need to drive across road divider in narrow roads, despite sufficient driving space (~07:17 mark, Tesla)

29. Detection of Construction Cones, Ex. 1: Unpredictable behavior with construction cones which are common on roads and roadsides (~02:30 mark, Tesla)

30. Detection of Construction Cones, Ex. 2: Unpredictable behavior with construction cones which are common on roads and roadsides (~08:15 mark, Tesla)

31. Detection of Construction Cones, Ex. 3: Unpredictable behavior with construction cones which are common on roads and roadsides (~02:44 mark, Tesla)

32. Detection/Avoidance of Barriers, Ex. 1: There have been reports of Teslas driving towards/into roadside barriers, here’s one example

33. Detection/Avoidance of Barriers, Ex. 2: An example of a Tesla on Autopilot coming into contact with a barrier on the left (~00:48 mark)

34. Detection/Avoidance of Barriers, Ex. 3: Here’s an ABC-7 news report on Tesla with Autopilot driving/crashing into barriers (warning: crash images)

35. Detection of Stationary Vehicles/Objects, Ex. 1: ADAS systems like Tesla sometimes have difficult detecting stationary objects right in front of them – here’s an example (warning: crash images)

36. Detection of Stationary Vehicles/Objects, Ex. 2: There have now been multiple reports of Tesla vehicles crashing into stationary police cars and firetrucks (warning: crash images)

torquenews.com/1083/police-te…

autonews.com/regulation-saf…

37. Object Detection (in Shade/Darkness): The difficulty that L2 systems face in detecting objects in darkness or shade is well known – here is an example (~3:24 mark, Tesla)

38. As an aside - it’s worth noting that real-world challenges faced by Autopilot are likely be similar with “FSD”

https://twitter.com/samabuelsamid/status/1321107947875946498

39. There are many more day-to-day, real-world examples where L2 ADAS systems need supervision or get disengaged – moral of the story is enjoy the enhanced safety and automation, but pay attention!

https://twitter.com/binarybits/status/1344738286955331584