Search results

It’s to give the rocket a nice steam bath, so it will be bright and shiny for the next launch.

That’s splitting hairs, but I’ll rephrase. The radar signature of a trailer and overhead sign may be identical. The Lidar signature is not. Thus, Lidar is far more useful than radar for a system that needs to be able to tell the difference, and act accordingly.

You seem to be confusing lidar with radar. See for example here: You don’t need to go autonomous to make trucking safer When the car is e.g. attempting to drive straight from a left-turn-only lane, or blowing through a stop sign, that’s a valid disengagement, whether the car would have crashed...

It doesn’t have to be 100% accurate/precise. Just accurate enough to distinguish a trailer from an overhead sign. That doesn’t require much accuracy, for a lidar. Truly you have a dizzying grasp of statistics. I’ve also occasionally had 10-mile drives with no disengagements. And 1-mile drives...

To radar, perhaps. Lidar, no. Lidar has much higher resolution than radar, and could trivially distinguish between these cases.

Sigh. The point is not that Waymo should change what they're doing, or that Lucids look better than Teslas. It's that Tesla could incorporate these sensors without compromising their aesthetics, which is what you seemed to be worried about. (Unless you were being sarcastic about that as well...

Of course not. But realize that THIS car has Lidar and multiple radars too (the Lidar is in the little rectangle just below the logo): Obviously Lucid's software to drive this sensor suite is not ready yet, but the point is that the hardware can be concealed; it doesn't have to be ugly.

I don't think anyone doubted (in the past 3-4 years anyway) that pure vision could achieve good-weather L2 with ~10 mile MTBF, which is where we are now. The skepticism was that it could scale to L4 with million-mile MTBF, which was Elon's promise from the beginning, and so far Tesla has not...

As I understand it, the IFT-1 flip maneuver was primarily required to keep Booster's propellant settled (by centrifugal force) while it separated from Starship, so it could safely start up its boostback burn a few seconds later. In the droneship scenario, with a Falcon-like separation, there is...

I don't work for Tesla, obviously. And I don't think they are about to announce this, though I think they have handicapped themselves tremendously by locking themselves into vision-only. I work in machine learning, so I have some pretty in-depth knowledge of all this. I can infer a lot from...

NN/training is the limiting factor as of today, with MTBF in the tens of miles. As the NN/training improves, hardware will become the limiting factor. For Robotaxi to be a success as Tesla envisions it, the ODD will have to be pretty wide. I don't think they can get there with pure vision. As...

Notice how they issued a software recall, not a hardware recall. The fault was with software, not hardware (radar/lidar/cameras).

Completely agreed; this is purely a NN/training failure, not a sensor suite limitation, and v12 currently has lots of such failures. True. But we are still orders of magnitude away from L4/L5. Once the low-hanging fruit is solved (largely NN/training), what remains will be largely cases that...

This is for the NN to figure out. As shown by the leap from v11 to v12, the NN can do a lot better job at this sort of thing than human coders can. If it turns out that the added sensors don’t add value (in terms of reliability), the NN will figure this out too, and will learn to ignore those...

10% relative to the wonky IFT1-style flip maneuver, but for a droneship landing scenario they could probably do a Falcon-type separation while flying faster and more horizontally, so presumably the penalty would be a lot less than 10%. By also eliminating the mass of the hot-staging ring, maybe...

The visualization network and control network are largely separate. That’s why the car sometimes seems to ignore signs (such as left-turn-only) that are obvious in the visualization. Even the best control networks won’t be of much use if you can’t properly see your surroundings. That’s where...

Lidar has every way to resolve this. It generates a 3D point cloud, which would look very different in the shadow case than in the obstacle case, whereas to the camera (or even to the human eye) the difference is far more subtle. The radar “saw” the side of the trailer, but had no way to...

The complexity of a large neural net can be surprisingly uncorrelated with the size of its inputs. For example, AlphaGo used a gigantic neural network with tiny inputs. Adding two sensors (radar+lidar) to the existing 7-camera HW4 suite (not counting cabin camera) wouldn’t increase the input...

I have two cars, both still on 3.25. (2017 M3 and 2022 MY.) Sent from my Job

Does the droneship capability imply landing legs, or would each droneship have its own catch tower? (I’m not sure how they could possibly do that, but fun to envision!) Also, without a boostback burn, would this eliminate the need for hot-staging (as it’s currently required to keep the...

It's not all beyond human abilities. As detailed in an earlier post, the camera sensors still have significant limitations relative to human vision. (Lower resolution, lower dynamic range, more vulnerable to glare/dirt/dust/raindrops, etc.) And L2 still includes emergency braking and collision...

I'm glad you're not programming the car! The job of the neural network is to provide a cohesive and consistent explanation for everything the sensors are "seeing". It will understand that if the camera sees something that might be a shadow but might be a pedestrian, that's an ambiguity, not a...

Part of this is limitations of the training set; clearly in good conditions the cameras can resolve the signs well enough to read them if the NN was trained to. And agreed that in good weather, perception is not the limiting factor. But when there are "Autopilot degraded due to poor conditions"...

But this is exactly how Tesla's eight-camera system currently works; it's like having eight separate viewers at once. The thing is that they all share the same "brain" (neural network), so the information flow between them is far more efficient and cohesive than a bunch of passengers trying to...

Yes. Hand-coding for sensors like radar/lidar is difficult because engineers have very little intuition for how to interpret the signals. But for a NN, it's just pure information. It will learn it like a bat learns to use ultrasound, or like a whale learns whale language (which humans have never...

Yes, hence the value of radar and lidar to improve the safety in such situations. The roads and lane markings and signals are pre-mapped, so GPS plus local positioning (being able to see lane lines within a few meters to gauge lane position) is sufficient for navigating those. That sort of...

That's very kind. I have two cars with FSD; a 2017 M3 and a 2022 MY. Both make mistakes with similar frequency on v12.3.6. So it is not the car's hardware. (Or if it is, it's a very common flaw, given that both of my cars exhibit it.) Likewise, unless I am actually hallucinating, the car...

Do you mean the strongest link? Neither radar nor Lidar have anything to do with sound; they are both photon-based. In any case, the "DOT mechanism" has much more to do with road rules, not with perception limitations. Obviously vision is necessary for any AV system. The point is that FSD-style...

The network is agnostic to the type of data. What it is exceptionally good at is resolving ambiguity and _apparently_ conflicting information. (There is only one ground-truth world outside the car, so there will always be a consistent explanation for any input; the NN just has to find it.) When...

I can't speak for spacecoin, but for myself on v12.3.6 I still have to disengage every 2-3 miles in city driving. Most of these are not safety critical, usually it's incorrect lane selection or driving on the shoulder or missing a turn or aiming for a pothole or being rude to other drivers or...

With hand-coded v11-style C++, you're right, it's difficult. One of the strengths of E2E neural networks (v12+) is that they can handle conflicting or fuzzy scenarios MUCH better than hand-engineered "if-then-else"-style coding can. But in a nutshell, the radar signal should be leaned on when...

Oh, absolutely. And there also should probably be a "Law of Regressions", which is that if you specifically train a neural network to solve a particular new set of scenarios, it's likely to cause it to regress in other areas. So 12.4 may have 5x fewer failures in the specific problem cases...

If the very first FSD release had had just 10% reliability, and then each successive version showed as much improvement as Elon claimed it did at the time, we'd be at [checks calculator] 99.999999999999999999999% reliability by now.

As pointed out, such HD maps are created locally with e.g. Lidar, not by satellite. In FSD's case, the maps can be helpful as a prior, just as your memory of a city layout is helpful when you drive it again. But they are not necessary, and FSD will override the maps when appropriate. (E.g. when...

Update: Tesla fixed it under warranty. Per the invoice, the fix was to "Install Occupant Classification System (OCS) Filter Module", and "Retrofit SBR ['Seat Belt Reminder'] in OCS Equipped Front Passenger Seat". I'm not sure if this was exactly the same issue covered by the recall, but in any...

The window really does close. (The required delta-V quickly exceeds Starship's capabilities.) Take a look at this plot: https://caseyhandmer.wordpress.com/wp-content/uploads/2019/10/explanation.png The blue regions are ideal, and anything outside blue-cyan is probably not realistic. A Starlink...

The low-energy transfer window is only a couple months long, every 26 months. My guess is that they might send several Starships per window with intentionally different designs, to learn as much as possible per window, and to have the highest chance that at least one will land successfully. If...

I interpreted "twice as strong" as toughness / fracture resistance. For the ablative layer, my first thought was that it might also function as an insulating layer, which would enable them to have the thinner outer layer of tiles. But I'm not sure if that's actually the case. Stainless steel...

Replying to myself, it seems like this two-layer heat shield structure is exactly what they’re planning to do:

Or half of the Texas Gigafactory.

220m diameter, not 2km. Still, huge! About 200 Olympic swimming pools’ worth.

This is my solution.

Tesla already has and uses astronomic amounts of training data. I’m not sure what you think would require it to be more astronomic than it already is? Granted it doesn’t yet know how to read no-right-on-red signs, but it will still make such turns safely, if not legally. Also note that by...

Elon begs to differ.

Undoubtedly it can do so 99.9% of the time. But can it do so 99.99999% of the time? That's the question.

Completely agreed that resolution is not the limiting factor. Dynamic range and low-light clarity are far more important, as well as positioning (bumper cameras) and resilience to dirt/rain obstruction. On the other hand, once the car is able to read and interpret nontrivial street signs, the...

Why do you say it has no contextual awareness? FSD is trained on video clips that are long enough to provide context, and certainly takes the last ~30 seconds into account when making decisions. The occupancy network is persistent, providing at least short-term "memory" of objects and places the...

AV's have the advantage of GPS and pre-mapped roads for much of this. Lane lines a couple meters away (to establish current exact position) are much easier to see in bad weather than cars or obstacles ~100 meters away which still might require pre-emptive braking; the latter is a huge advantage...

Lucid has a ton of sensors, but not all of them are activated. I was also surprised to learn (at a showroom recently) that most of their fleet actually doesn't have Lidar installed. I would have expected they would all have Lidar just to gather fleet data like Tesla does, but apparently not...