Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Tesla has the largest sensor fleet on the road [1] (~370k vehicles with AP2.0+ hardware, adding ~25k units a month to that) contributing to the refinement of their models (~1.2 billion miles of shadow mode experience). Waymo may be closer because of Google/Alphabet resources, but I can't buy a Waymo vehicle nor ride in one, and I'd rather give Tesla the money than Google.

[1] https://i.redd.it/qvrw7zcvw8821.png



I got my S90D 3 weeks before AP2 came out, and I was told it was not backwards compatible. At the time, it was said that this upgrade was required for FSD.

I swapped in October to a 3PE, now I'm seeing that another upgrade is incoming, and that THIS one is required for FSD.

I've spent plenty of money with this company, but handing over $5,000 more dollars for a feature that is unproven is not something I feel like I should need to do; especially when I'm clearly advertised to at the time that I can buy the feature later for $6,000.

If fully autonomous driving becomes available, and it costs more than what was advertised, they are going to have a big problem.


I doubt, and I say this as a computer vision researcher, the problem will be solved (if it is solved) by just a camera array.


I am surprised, given that humans seem to drive mostly by a pair of video inputs.

Can you go into depth about why this comparison is inappropriate, why it unlikely to be solvable in this way?

(I can certainly understand “not now”, but your claim seems rather stronger than that).


It’s really simple: our video inputs are better than cameras, react in a wide range of light conditions, and most of all are hooked up to a GGI: Genuine General Intelligence, aka human brain. Even then it takes us almost two decades before we’re safe to drive, and we still smash into things.


Unlike the Scarecrow, we also use our brains!

We’ve made extraordinary progress, but, ultimately, we’re still training systems in extremely simplistic ways to solve (conceptually) simplistic tasks. Localizing objects (into a set of categories no less) is only a fraction of what your brain does when looking around. Moreover, we still struggle with this. One of the Tesla accidents was due to a segmentation error (two simimilary colored and textured objects were merged).


If I were investing in this area - I'd buy shares in makers of highway striping paint and firms that make snow-clearing equipment.


Except it doesn't work where it refuses to work and the cars don't actually send back any meaningful amount of data (unless you want to train your model on lossy compression algorithms).

You'd need a storage rack in your frunk to record any amount of raw sensor data. That's before you consider their production AP hardware doesn't have cycles to spare to do any of that recording.


> the cars don't actually send back any meaningful amount of data

Source?

> You'd need a storage rack in your frunk to record any amount of raw sensor data. That's before you consider their production AP hardware doesn't have cycles to spare to do any of that recording.

I have no idea on what you base your claims, but without more information it is hard to figure out what you have in mind...


Do the math. 7 cameras at 1280x960 with 16 bpp at maybe an average 30 FPS and you're talking 500 MiB/s of just raw image data you would need to record. Where is it going? Not on any sort of storage and not over the wire to Tesla, that is clear.

What curious people have found is that the car just sends a basic disengage report with maybe 10s of h264 video or spaced far apart raw images from the cameras.


Who the hell stores or sends raw image data? The training data that's used for offline training of the vision neural nets is probably all h264 encoded anyway.


Thanks.

You don't (always) need full resolution with 30 fps raw frames for training your model though. If you are looking for missing exit, stopping at a traffic light, etc. you need front facing camera (so 3) and a few raw frames (and the sensor is 12 bits not 16). One frame is 1.8MB raw, but if you train on YUV images, this would be even smaller. With a 200MB buffer in memory, you can keep 100 frames at any time (and you have 8GB local storage for storing while waiting to upload)

You said "any meaningful amount of data", I guess this is subjective!


Something that hasn’t been true for some time. Some nights my Model 3 will upload 4-5G of data. What do you think that data is Einstein? My Easter egg game high scores?

https://electrek.co/2017/06/14/tesla-autopilot-data-floodgat...


Can those sensors recognize traffic lights?


They can't distinguish billboards from real threats so the promise of FSD is complete vapor.


Can cameras not detect traffic lights and their signal colors, assisted with mapping data to give a fuzzy idea where traffic light recognition will be required?


Not sure if I would rely on mapping data. There are so many new-ish roads and lights around my apartment. After living here for 6 months one of the roads still isn't known by Google Maps, Apple Maps, or my Tesla's nav software.


If they relied on mapping data to help them know where to look for the lights, I wonder how they'd cope with temporary signals (e.g. at road works)?


You can't drive using maps.


Doesn't Cadillac SuperCruise still outrate Tesla in terms of its refinement?


Wasn't that the article about how it was better at disengaging than autopilot?


I have yet to see any evidence or proof of this, but am happy to be proven wrong if someone can provide a legitimate citation this is the case.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: