Hacker News new | ask | show | jobs
by martin_bech 3507 days ago
In my autopilot gen 1 Tesla, when the software was first released, lane markings had to be really clear, but as the software has matured, it now does less than perfect lane marking in pretty heavy rain without issue.
2 comments

What does it do on roads without any lane-markings? My country has lot's of them outside the cities.
It does two things.

Holistic path planning uses the entire scene to identify the lane, not just the lane markings.[1] This is exactly how a human drives on a rural road.

Once enough fleet learning data is gathered, high resolution maps will mark out the road surface to high accuracy (cm).

Of course that's no good unless the car knows where it is within cm! For this Tesla uses sensor fusion with GPS, an onboard IMU (similar to a smartphone), four wheel wheel odometry (via the ABS sensors), steering angle, and torque delivery. Owners report that the Autopilot 1.0 cars know their position in a parking space to within about 10 cm. And it has demonstrated high accuracy and low drift in no-GPS situations[2].

This type of sensor fusion (Kalman filtering) was originally developed for the Apollo spacecraft, and was one of the responsibilities of the famed Apollo Guidance Computer.

[1] http://wccftech.com/tesla-autopilot-story-in-depth-technolog...

[2] http://motherboard.vice.com/read/how-teslas-model-s-tracks-y...

I wonder how it works in situations like we have in NZ where the recent earthquakes have moved parts of the country six metres.
Or heavy rain will do the same. When it's dark and rainy, we often have to drive almost blind here.
I think the current-gen autopilot vision code is a completely different codebase from a different supplier running on different hardware than the gen 1, though, so all those improvements have probably been thrown away.
Despite the different codebases, they still have the datasets that they've collected over the years, and that's really what matters (for things like lane detection.)

It would be quite weird if they didn't retrain their new models with all that data.

What does this mean? So they send all 8 video cams data to Telsa? This would mean ould mean very high upload traffic and costs - sounds unrealistic. I assume Tesa sends the GPS data, and sensor data and maybe occasional photos back to HQ, but not full videos of all 8 cameras - or do you have a source/reference?
Of course they don't send all that back, but they didn't train the original on that data either.
Curious, why do you think so? Wouldn't make much sense to throw all that stuff away.
For the same reason it makes no sense to keep your Sega Saturn games when you now have a Playstation.

There are some testing scenarios which Tesla should be able to use, ut none of the software from MonilEye has relevance to Nvidia's platform.

Actually, i think in AP 1 they used MobileEye and Tesla has their custom software to compliment it and with the switch to AP 2 "It runs a Tesla-developed neural net for vision, sonar, and radar processing." https://cleantechnica.com/2016/10/25/inside-nvidias-new-self...

I also read another quote from the Tesla engineering team that "nearly all of it" (the data collected from AP 1) can be used for the new AP system https://electrek.co/2016/11/07/tesla-bought-1-million-data-s...

So it seems like the analogy is more akin to Xbox 360 and Xbox One, hardware got an upgrade but they were able to run Xbox 360 games on it with some work.

I know I'm probably gonna get downvoted, but I cannot not bite. I'm super happy I kept my Saturn games. I still play them occasionally and some of them are very valuable now. I also hope you meant PS3 or PS4 when you said "PlayStation" and were not trying to revive a long gone console war that was very painful for some of us.