Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

FSD Beta Videos (and questions for FSD Beta drivers)

This site may earn commission on affiliate links.
Most likely you are using software that works through standard ethernet equipment, in which case there's a lot of components (including sometimes the software itself) in between that can lead to dropped frames unrelated to the system being wireless (AKA it'll do the same even if the connection was wired). You can analyze the wireless connection for dropped packets (easy way to do this is just send a ton of pings, but there are other tools that can do so also). It's easy to find examples of people using wired connections with dropped frames when using NDI or NDI HX (lots if you just google it).
I use NDI|HX. That's what breaks down over 802.11ac or faster Wi-Fi. Hardwired networking is the only viable approach, even with a completely 100% dedicated Wi-Fi network in a building with no other Wi-Fi networks far away from anything. And wired networks are flawless with the same hardware on both ends.

Pings are a bad way to analyze a network, because they have the lowest priority of any traffic. QoS on the router or switch will drop them on the floor far more often than real traffic.

Either way, NDI|HX latency is about an eighth of a second (even over wired Ethernet, and IIRC slightly higher over Wi-Fi). Any compressor technology would have to be bespoke to operate with low enough latency to be useful for autonomous driving.
 
#FSDBeta v9.2 Turn Testing and Acceleration Analysis

Second video today, check it out.

It really seems like they haven't trained the neural net on oncoming cars more than 200ft away.
I would expect there to be a huge improvement in this behavior at some point soon...
At 6:06:
1629044658400.png
 
  • Disagree
Reactions: WattPwr
Elon dropped the technical notes for V9.2


Finally, Tesla gives us the technical release notes!!! I am very excited!! This gives us real insight into what Tesla is doing with FSD. The release notes show that Tesla is working on prediction and planning!!
That's a major and very welcome change in terms of transparency. Some of the language is above my pay grade and I hope that the knowledgeable among us will decipher both the meaning and the impact.
 
Elon dropped the technical notes for V9.2


Finally, Tesla gives us the technical release notes!!! I am very excited!! This gives us real insight into what Tesla is doing with FSD. The release notes show that Tesla is working on prediction and planning!!
I believe VRU = vulnerable road user (pedestrians, cyclists, motorcycles)
 
I wonder if Tesla will be talking about some of this as part of AI day, so maybe that's why there's some increased transparency. But it can also show how they are behind in some aspects, e.g., this behavior is "new" or "v1?"

Multi-modal predictions might suggest Autopilot has been making decisions based on the most-likely behavior of other vehicles so far, but other vehicles have some probability to turn or change lanes.

I'm hoping the "Lanes network" is reducing the reliance on correct "lanes" data from OpenStreetMap, which sometimes has the wrong / missing number of lanes for a direction or turn. This has previously resulted in sudden merges across intersections where it assumes the other side only has a single destination lane.

As others have pointed out, VRU is likely referring to pedestrians and bikers, and my guess of "new crossing/merging targets network" could be referring to better fusion across cameras for pedestrians that move from pillar camera to fisheye. Although it could be literally crossing as in cross-walk.

Quantization-Aware-Training seems to be improving accuracy / working around int8 only storing 256 possible values, so rounding errors when converting for something that expects more granular floating point values could result in strange behavior, e.g., things jumping around.
 
It's a huge improvement over the 9.1 video.
Where are you seeing it exactly? The usual turn looks basically identical and I'm pretty sure the other lefts would have been identical before this update, we already know the system handles particular lefts differently based on visibility or whatever other factors

But the usual left that is in so many of Chuck's videos, the behavior looks identical to me
 
Last edited by a moderator:
  • Like
Reactions: Matias
Where are you seeing it exactly? The usual turn looks basically identical and I'm pretty sure the other lefts would have been identical before this update, we already know the system handles particular lefts differently based on visibility or whatever other factors

But the usual left that is in so many of Chuck's videos, the behavior looks identical to me

Check out the forward facing lefts compared to this one with 9.1, where it was making dangerous decisions to proceed with oncoming cars:

9.1 video below, not 9.2

In the 9.2 video, it was essentially making all the forward lefts safety, 100%
 
I use NDI|HX. That's what breaks down over 802.11ac or faster Wi-Fi. Hardwired networking is the only viable approach, even with a completely 100% dedicated Wi-Fi network in a building with no other Wi-Fi networks far away from anything. And wired networks are flawless with the same hardware on both ends.

Pings are a bad way to analyze a network, because they have the lowest priority of any traffic. QoS on the router or switch will drop them on the floor far more often than real traffic.

Either way, NDI|HX latency is about an eighth of a second (even over wired Ethernet, and IIRC slightly higher over Wi-Fi). Any compressor technology would have to be bespoke to operate with low enough latency to be useful for autonomous driving.
for reliable automotive-grade networking, the future (imho) is with AVC (which is the old name; audio video bridging) and the new name is TSN (time sensitive networking).

to play in the TSN game, you need a whole bunch of new networking goodies. ptp for timing (iee1588), high-end gps and special NIC hardware that squirts the NIC-based RTC timestamp into the outgoing packets just in bit-time as it exits. its really cool stuff and its getting there, for cars; but the hardware is still a bit expensive, the switches all have to be tsn-aware, the software and config is kind of new to most vendors and so on. but once it is understood and implemented by the vendors, then we'll have level-5 capable infra that we can really count on.

you can get experimental TSN stacks for host-based linux, and some autosar companies also are starting to offer it.

wireless (wifi) is useless. give it up, guys. wifi is so far behind in all that car networking needs; the only time you use wireless is for end-user stuff or low reliability backwards compat use-cases. lte is for WAN use, and that has nothing in common with wired ethernet, since the reasons for them existing and their speed needs are quite different.

for high speed reliable networking, its not only *wired* but the new kind of wires (including broad-r-reach gig-e and redundant switch ports).

oh, and pings can be useful for testing; but you do have to run qos and assign them prio's so that they dont get dropped.

use udp based tools (netperf2 is still one of my favorites) and you can see the latency, jitter, frame-loss and thruput. its not a Spirent (I love those) but its also not a $100k bit of kit. regular raspi pi4 can fill a gig-e pipe with iperf reliabily.

take it from someone who does this stuff.
 
Is your mind boggled easily? Or what exactly do you find mind boggling?
I can't speak for anyone else, but my mind is definitely boggled by V9.2

Not for anything in any of the videos as I haven't watched them, but actual FSD release notes.

Actual release notes from Elon? Did our reality just break?
 
I wonder if Tesla will be talking about some of this as part of AI day, so maybe that's why there's some increased transparency. But it can also show how they are behind in some aspects, e.g., this behavior is "new" or "v1?"

Multi-modal predictions might suggest Autopilot has been making decisions based on the most-likely behavior of other vehicles so far, but other vehicles have some probability to turn or change lanes.

I'm hoping the "Lanes network" is reducing the reliance on correct "lanes" data from OpenStreetMap, which sometimes has the wrong / missing number of lanes for a direction or turn. This has previously resulted in sudden merges across intersections where it assumes the other side only has a single destination lane.

As others have pointed out, VRU is likely referring to pedestrians and bikers, and my guess of "new crossing/merging targets network" could be referring to better fusion across cameras for pedestrians that move from pillar camera to fisheye. Although it could be literally crossing as in cross-walk.

Quantization-Aware-Training seems to be improving accuracy / working around int8 only storing 256 possible values, so rounding errors when converting for something that expects more granular floating point values could result in strange behavior, e.g., things jumping around.
As i said in the other thread:

Elon just posted that they are at their first version of multi model prediction and its not even fully implemented yet.
In contrast, Waymo has been doing its for years. Infact they have moved forward from there to better multi-modal models to the point they released a paper about their previous multi model architecture in 2019 which was SOTA.
https://arxiv.org/pdf/1910.05449.pdf

So while @powertoold @rxlawdude has been claiming Tesla supremacy. The reverse has been true time and time again. When the truth comes out we find out that Tesla is actually years behind in NN development and deployment for AV. This is truly mind-boggling!

This is backed up by Green's analysis of the prediction models FSD Beta has been using.

EsqwfcaVkAACmoS
 
Last edited:
  • Like
Reactions: diplomat33