Comment by TooKool4This | Hacker News Reader

Comment by TooKool4This | Hacker News Reader

This video shows a lot of other ,,interesting" issues with the current FSD implementation. The collision with the pylon is in fact the least interesting thing.

- 2.20: Taking a left turn with a trajectory into the opposite lane. The youtuber's camera clearly was adjusting exposure from underneath the overpass and I wonder if the car's cameras had the same issues.

- 2.40: Right on red without stopping. I wonder if this is related to the stop sign rolling stop issues

- 3.30: Collision with pylon. It seems like the perception system didn't pick up the pylon at all (or if it did, its not rendering on the screen)

- 5.10: Some weird funky trajectory and then pulling into the sunken curb/sidewalk on the right turn

- 6.28: Trying to drive down the railroad tracks

I gave up after the above.

Despite Tesla's (Elon) insistence that FSD will be safer than human drivers soon enough (1 year), there is very little evidence to support that. At the very least if we are going to allow beta testing on roads then Tesla should be forced to submit their FSD incident data to the CA DMV like the rest of car manufacturers that are testing AVs.

2:20 is wild -- the street it's turning onto is clearly marked as a two-way in Google Maps [1] (which I understand provides the underlying data to Teslas). Why would the car think, no, it's best I turn onto the left side of this road?

I'm also in awe of how fickle its predicted path is, especially when turning. Why commit to a new path for a tenth of a second and then change its mind again? This suggests to me it's placing way too high weight on poor quality telemetry. Heck, at 5:14 it can't even decide which street it wants to go down. Maybe it's just a UI thing (always displays the most likely path) and the underlying model is actually keeping both options open.

At 6:45 it tries some insane passing maneuver -- I can't get over that one. Passing on the right, in a bus-only lane, planning to squeeze through the rapidly-closing gap of the car in front of it and a parked car, where there is currently a bicyclist. And it looks like the bicyclist is the only thing keeping it from doing that insane maneuver -- each time the bicyclist is occluded by the car in front (object permanence, hello?), it attempts to pass.

I've lost count of the number of things this car doesn't understand, that it should to be on a road:

* object permanence

* reading basic text in simple print

* logical connections between the map and road geometry

* patience and commitment

* understanding physics and geometry of other moving and non-moving vehicles

* seeing stationary bright orange objects immediately in front of it

[1] https://goo.gl/maps/EQfUnPrN52d9TVKPA

AFAIK Tesla doesn't rely on map data for the driving itself, only for routing. Which sounds like the right approach considering roadworks and how often maps are out of date.

Yes, this is very far from average human level performance. Error rates need to go down by maybe 100-1000x to be viable. They are years away.

As a foreigner, can you explain at 2:40, which light is red? I can only see a green light on the right side, there was no traffic light on the road it came through. Also doesn't the US allow right turns on red?

I would totally drive on those train tracks myself.. it's pretty common here, and there was no signage. Curious to see what the AI would do next if the driver hadn't intervened.

You still have to come to a stop before turning right on red.

> - 3.30: Collision with pylon. It seems like the perception system didn't pick up the pylon at all (or if it did, its not rendering on the screen)

Wonder if LIDAR would have spotted the pylon? It sounds like neat tech.

> that FSD will be safer than human drivers soon enough (1 year)

maybe while parked…

Wow, thanks for the timestamps. This wouldn't even pass a driving test.

Reply on news.ycombinator.com