Hacker News Reader

Comment by sxp | Hacker News Reader

From http://www.teslamotors.com/models/design

>Model S comes standard with a forward looking camera, radar, and 360-degree ultrasonic sensors that actively monitor the surrounding roadway. Progressive software updates over time will enable sophisticated convenience and safety features that use these sensors to respond to real world conditions. These features will ultimately give Model S Autopilot capability on the highway from on-ramp to off-ramp.

That's a pretty nice TODO to leave in the code :)

View on HN

At this stage all you need to install is the machine learning algorithm so it can watch and learn what real drivers do when faced with the conditions it's sensing.

Absolutely brilliant!

Machine learning can't do that and won't be able to for a long time even in principle, to say nothing of doing it in real-time on a car. Autonomous driving contains very little machine learning for all the sensing and high-level parts.

It may contain some low-level classifiers for things like people and possibly other cars, but this is not strictly required and car manufacturers generally try to avoid it, because it is difficult to reason about. Traffic sign classification is another candidate, but again, at least the Google car probably pulls this from their precomputed map.

Finally, none of these classifiers are trained online, in the car, but always offline, likely on carefully selected data on some GPUs for many days or weeks.

Fire-and-forget machine learning may be a bit overexuberant, but one should not discount the value of these data. On one hand, there is sensory data about the real road environments these Teslas are traversing. On the other hand, we have human driving decisions. Together, one has–at at a minimum–a rich test environment. At a stretch, it could be analogous to the corpus on which Google developed its translation capabilities.

The nightly reporting-home of Tesla's fleet shows how interconnectedness is introducing network effects into unexpected places. Sort of like how map quality is now driven–in part–by the user base.

By this time next year when they have 40,000+ vehicles all logging conditions and driver decisions, in car processing and OTA updates to their testing algorithms may be the better option. Google's fleet come back to base every day, Tesla have to think very differently in their approach and bandwidth could be their biggest issue.

Actually, machine learning can do that. Back in 1995, a team at Carnegie Mellon used a neural network in a car:

"ALVINN is a forward-looking, vision based driving system that uses a neural network to learn the mapping between road images and appropriate vehicle turn radius. By watching a person drive for about 5 minutes, it can learn the relevant features required for driving[3]. It has successfully driven our testbed vehicles on unlined paved paths, jeep trails, lined city streets and interstate highways."

http://www.cs.cmu.edu/afs/cs/usr/tjochem/www/nhaa/navlab5_de...

Wouldn't recaptcha video would help google?

The problem obviously is ... Privacy and more importantly - upload speeds to google servers. US internet infrastructure wouldn't cope with thousands of cars uploading crisp HD videos. Probably google infrastructure wouldnt cope either.

Meh, there's no way their hardware is equivalent to what Google or the like put on their self driving cars. So I don't think any algorithm can realistically learn everything based on highly limited information. I feel like it's much more likely to enable some safety functionality than anything approaching real autonomy.

Their hardware is almost identical to Google's, except for the cloud point laser scanner on top of the vehicle.

For the Google car, LIDAR is the absolute primary source of accurate data. It's like saying that a car without an engine has almost identical hardware as it has everything else.

Can you describe what can't be detected without a LIDAR scanner and without the Tesla's onboard ultrasound sensors?

@toomuchtodo: For the Google vehicles, LIDAR is the primary sensor. They're not identical in the slightest.

@agildehaus

Elon and Tesla believe that their array of ultrasound sensors, GPS, and front facing camera will be sufficient without the addition of an expensive LIDAR roof mounted scanner. Google has under 100 self-driving vehicles in the field collecting data; Tesla is now making 2500 vehicles/week that will be collecting data.

My comment stands. The sensor networks are the same except for the Velodyne LIDAR device. Google has Street View data, Tesla will have a substantial fleet of vehicles collecting data towards improving their autopilot system.

Vast amount of mapping data do you no good if you have to test your scenarios in simulated environments (as Google requested they be allowed to do to the California DMV).

https://www.youtube.com/watch?v=g3WtvzmKCQQ

I don't think 'our system has been trained to respond in the same way as the average driver' is much of a selling point. You want your auto drive to do the right thing, not the average thing.

most drivers do the right thing most of the time. Accidents are usually caused by the occasional mistake from humans being unreliable.

Reliably driving like the average driver is perfectly acceptable.

(On a side note, nothing is preventing the machine learning from learning when a driver's response results in accident, a system can learn from negative experiences as well)

Reply on news.ycombinator.com