Mobile eye CES2017 presentation and Teslas Autopilot future?

[Ragnarr]

Hey guys,

So, I was watching the Mobile eye CES2017 presentation, and I'm starting to get a little worried about the state of Teslas level 3 & 4 autopilot progress.

In the presentation, it was made very clear that it's important to have a front and rear lidar systems and we all know Elon doesn't like using Lidar.

If I understand, correctly Audi A8 has this level 3 system in their new vehicles. (It operates only in a traffic jam as for now) and BMW is testing and plan to release a car in 2020.

The mobile eye is taking the same eight camera approach as Tesla except for using Lidar for redundancy.

I know their REM system is kind of similar to what Tesla is doing by crowdsourcing and collecting data and building map data from existing vehicles.

My Model X is coming in May, and by all means, I want Tesla to win. But given the lack of updates and "coming soon" feature promises do you think Tesla can solve the problem without using Lidars? I know the weather has an impact on the lidars and when I test drove the MX it was the cold, rainy and dark weather but the autopilot worked flawlessly. But I want Tesla to win the Level 3 or 4 game before others.

From their Level 3/4 demo, it seems Tesla can do the full autonomous driving with just cameras and radar but whats the holdup? Do they have to collect more data?

Where I live (Norway), we have too many Teslas, and I'm sure there is a tonne of data already.

Any sort of technical input would be great.

 

[Derek Kessler]

Do you, dear human, have Lidar?

 

[Gwgan]

we have too many Teslas,

No such thing.
The Tesla Show – Podcast Episodes Has some good color on this very topic, with an update at the end of the one on this recent conf call.

 

[Snowstorm]

Do you, dear human, have Lidar?

No, but I do have 2 amazing camera with crazy dynamic range, mounted on a 3 axis pivot, all connected to a super duper computer trained with millions of year of natural selection. Oh there are tactile accelerometer, and stereo microphones too.

I do get tired, get distracted and make random acts of poor judement though...

 

[jelloslug]

The Audi system that has been demonstrated is almost as good as the 2014 AP system on expressways only.

 

[NeverFollow]

So, I was watching the Mobile eye CES2017 presentation, and I'm starting to get a little worried
about the state of Teslas level 3 & 4 autopilot progress.

Lidar or not, autopilot should always provide enough braking distance to avoid unexpected obstacles.

In the case of the Tesla hitting a firetruck on a freeway, the car in front of the Tesla noticed the firetruck and changed lane.

One of my big worry would be that while in a curving road, an autopilot would not be able to notice and avoid a pedestrian or a bicycle.

 

[cwerdna]

From their Level 3/4 demo, it seems Tesla can do the full autonomous driving with just cameras and radar but whats the holdup? Do they have to collect more data?

Perhaps you should follow the thread Autonomous Car Progress.

Be sure to pay attention to stuff like Autonomous Car Progress, the 2015 thru 2017 California DMV autonomous vehicle disengagement reports (linked to from Testing of Autonomous Vehicles for Tesla vs say Waymo or more recently, Cruise Automation) along w/other efforts other players have made (e.g. Nissan and Cruise Automation taking along jouranlists, Waymo's service in Phoenix, etc.) Why testing self-driving cars in SF is challenging but necessary is interesting and IMHO, supposedly running in shadow mode with no actuation of a vehicle's controls and no resulting reactions from others that share the road is insufficient. Tesla's clearly not doing this type of testing the city of SF (under 65 km away from Tesla's HQ) or any public roads in CA, at least not between 2015 and 2017.

And remember, Silicon Valley where Tesla is headquartered is a hotbed of software engineering talent.

Do notice the dates of
CA DMV Report Sheds New Light On Misleading Tesla Autonomous Drive Video - DailyKanban (even if you totally disagree with the allegations and conclusions) vs. https://www.dmv.ca.gov/portal/wcm/c...a/Tesla_disengage_report_2016.pdf?MOD=AJPERES.

 

[cwerdna]

Lidar or not, autopilot should always provide enough braking distance to avoid unexpected obstacles.

Is that really realistic? And would that yield an autopilot that would yield acceptable behavior for the driver?

A quick Google search turned up 2017 Tesla Model S P100D First Test: A New Record - 0-60 MPH in 2.28 Seconds! - Motor Trend, mentioning 60 mph to 0 mph in 109 feet, which is on the pretty good side.

The best contenders at Best Car Safety Performance ranged from 107 to 119 feet.

Model S is 196 inches long or 16.33 feet. 109 feet implies needing to maintain at least 6.67 Model S-lengths between the AP driven car and the one in front.

On busy LA highways moving at the speed limit or higher, that sort of distance is probably not realistic or acceptable to the "driver" at all times. You'd have people cutting in front of you a lot.

Not sure about where you drive since your location is unknown...

 

[Ragnarr]

Lidar or not, autopilot should always provide enough braking distance to avoid unexpected obstacles.

In the case of the Tesla hitting a firetruck on a freeway, the car in front of the Tesla noticed the firetruck and changed lane.

One of my big worry would be that while in a curving road, an autopilot would not be able to notice and avoid a pedestrian or a bicycle.

Funny you said that my wife's VW egolf has adaptive cruise control. I was using it on a 50kmph road and approached a roundabout. There was a vehicle in front of me and I kept my hands on the steering wheels and let the radars control the speed. Big mistake. As soon as the car in front went out of radars visibility while making the turn my car suddenly speed up almost going off the road. I managed to hit the breaks. So yeah I think it's the same situation here.

We know AP 2 really messed up in the latest crash and unfortunately killing the driver.

I just think Tesla needs more redundancy and especially Lidars. Look at the Waymo's test drives and you will see the lidar cloud giving a nice situational awareness at 360' and keeping the cars safe.

I just hope Tesla has better reasons than saying Lidar won't work in bad weather for not having additional redundancy especially in situations where human lives are at stake.

 

[Snowstorm]

I just hope Tesla has better reasons than saying Lidar won't work in bad weather for not having additional redundancy especially in situations where human lives are at stake.

This isn’t a problem sensors (Lidar or something else) will necessary solve. The camera can perfectly see the barrier. It is the processing of the signal that needs to be improved so AP can understand that there is a barrier that needs to be avoided.

 

[Ragnarr]

This isn’t a problem sensors (Lidar or something else) will necessary solve. The camera can perfectly see the barrier. It is the processing of the signal that needs to be improved so AP can understand that there is a barrier that needs to be avoided.

But, That would take a lot of training before the NN learns various shapes and obstacles? From my understanding Lidar seem to give a better 3D map of the surrounding and maybe they can combine the data from lidars with what camera sees to make better decisions.

 

[Jjadmd]

How did the Lidar work out for Uber? Just saying...

 

[Snowstorm]

Better sensors may make processing easier, but processing is still the key lacking piece right now.

 

[jeewee3000]

But, That would take a lot of training before the NN learns various shapes and obstacles? From my understanding Lidar seem to give a better 3D map of the surrounding and maybe they can combine the data from lidars with what camera sees to make better decisions.

Of course adding LIDAR to the current AP2 hardware would improve data collection and redundancy, but the assumption that LIDAR is crucial for autonomy is what Musk disagrees with.

The recent Model X crash was - as has been said here - not caused by the car not seeing the barrier. It has three forward looking instruments:
1) Main Forward Camera - max distance 150 meters;
2) Narrow Forward Camera - max distance 250 meters;
3) Radar (only forward facing) - max distance 160 meters.

Tesla's current AP software is not yet reacting correctly to its different input signals. The needed progress therefore lies in the software. How fast Tesla will achieve level 3/4/5 will depend on their ability to develop this software.

I am not saying Tesla will win the race to self-driving, but they will most likely win the race to having affordable self-driving in a consumer car such as the Model 3 (and of course S/X/Semi/maybe Roadster 2.0/Y/...).

If you are worried Tesla will fail without LIDAR, I would take solace from Elon's recent statements regarding how amazing their in-house AP is already. It just cannot be released yet until it is more proven/polished. Have faith.