Hacker News Reader

Comment by Fricken | Hacker News Reader

Here's the slideshow that accompanied GM's presentation to investors yesterday. It's a breezy read:

https://www.gm.com/content/dam/gm/events/docs/5265893-685163...

Cruise CEO Kyle Vogt, in his part of the talk repeatedly emphasized the efficiency of their validation process that contributes to their high rate of improvement. Every outside indicator suggests Cruise really is moving faster toward their goal than anyone else.

Vogt also repeatedly emphasized that they have a mountain of engineering challenges still ahead of them before they're ready for an initial commercial deployment. They're rapidly scaling up testing, and hope to drive 1 million miles/month by the end of 2018. For Comparison, Waymo just announced they've done 4 million miles testing on public roads total, since 2009.

View on HN

From Waymo's 4 million mile announcement[0], they did their most recent million miles in 6 months (May 2017 to Nov 2017). The announcement also has the timeline of when they hit each million mile mark.

Also, I'm not sure this stat says anything besides how many cars they have on the road. Disengagement rates are probably still the most useful to gauge how good any company is doing.

[0] https://medium.com/waymo/waymos-fleet-reaches-4-million-self...

Disengagement rates out of context don't indicate much. Going 10,000 miles up and down a lonely stretch of highway without an unplanned disengagement makes for a pretty number, but is worth less than going a mile in dense urban traffic with cyclists, pedestrians, left turns through busy uncontrolled intersections, double parked vehicles etc.

Is there a link to the talk?

Yes there is. click on the 'webcast' button, top right:

https://www.gm.com/events/5265893.html

oh, they're lidar based. those sensors are super costly.

FTA: "GM says the cost of the next-generation lidar will be cut by about half to $10,000."

That's a misquote. Lidar is down to $10k from $90k. With the acquisition of Solid state lidar startup Strobe, they hope to get the costs down to ~$300. It's not clear if that's a per-unit cost, or a per-vehicle cost.

For sensors Cruise's Bolts currently use 14 cameras, 3 articulating radars, 8 regular radars, 5 Velodyne32 ultrapuck Lidar units, and 10 ultrasonic proximity sensors.

5 lidars! wow

That's what Waymo's little bug-like car had. One in front, one in back, one on top, and one below each side-view mirror. Full coverage.

LIDARs are getting cheaper. Continental, the big auto parts maker, bought Advanced Scientific Concepts and is preparing to bang out low-cost LIDAR units in quantity.

Color TV cameras were once insanely expensive. This is the RCA TK-41, the first good color TV camera.[1] Each image orthicon alone (it used three) was about $10,000 in the 1950s. The camera in your phone costs about $10 and far outperforms a TK-41.

[1] http://www.earlytelevision.org/rca_tk-41.html

Flash LIDARs have some serious problems:

They require extremely high power outputs since the beam spreads out. This limits the range significantly but also makes it so that you need very powerful lasers, and the only way to make that eye safe is to use 1550nm light, which requires InGaAs photodetectors, which are much more expensive. In addition, you usually need SPADs which have very high noise levels. You can reduce this noise somewhat by cooling them, but that's very expensive and infeasible for automotive applications m

That's what I thought when I visited Advanced Scientific Concepts in 2003 and saw the prototype optics working on an optical bench. They pointed the thing out an overhead door and imaged the parking lot. It worked. But it was too big and too expensive for the 2004-2005 DARPA Grand Challenge.

Over time, ASC units got smaller, range and resolution improved, but pricing remained around $100K. Then Continental, the European auto parts company, bought ASC. Continental's units are smaller, cheaper and more rugged.[1] They haven't announced a price point yet, but being a major automotive parts manufacturer, they know how to get the price down on something when they make it in volume.

Continental is quietly coming out with all the parts for self-driving cars. They make the sensors, actuators, and high-reliability computers. They work on the little stuff, too, such as systems for cleaning the sensors while in use. All the things you need to do it, instead of just prototype it.

[1] https://www.continental-automotive.com/en-gl/Landing-Pages/C...

Again, it's not like they don't work, but the point is that it's hard to shrink and make cheaper while retaining good performance.

They acquired lidar company Strobe in October: https://medium.com/kylevogt/how-were-solving-the-lidar-probl...

Reply on news.ycombinator.com