Hacker News Reader

Comment by Joel_Mckay | Hacker News Reader

Camera based guidance systems are unreliable, and the number of edge-case failure modes grow exponential in time.

LIDAR/LADAR based systems are not perfect, but do offer mm precision for guidance systems. SLAM based LIDAR systems can be very good, but are also not perfect when forced to guess where a platform is located.

Cheers, =3

View on HN

> edge-case failure modes grow exponential in time.

How so? Honestly asking.-

That is a long explanation, but generally even human binocular disparity incorrectly guesses 3D structures and distances. Our brains automatically fill in a lot of missing information that computers just can't know a priori (example: you know a dog hidden behind a car doesn't actually vanish nor remain stationary.)

Most guidance platforms would use LIDAR/SLAM to describe the local road surface, and overlap camera vision data to extrapolate distant surfaces and objects. Note distant objects also have lower resolution, unknown non-distinctive features (speed bump, or open man-hole cover etc.), and increasing sparse data as velocity effectively lowers world-state sampling rates.

The world-state is constantly changing at every intersection, sampling constraints add latency, and the navigation way-point goals may reach contradiction with immediate path-planning due to ambiguous/expired information.

Cheers =3

So, it doesn't "grow exponentially", you just wanted to express that there are "many".

> you know a dog hidden behind a car doesn't actually vanish nor remain stationary.

Computers can do that, too. It's not that different ... guessing missing words.

I think it was two years ago that Tesla told the public that they could do that.

>Tesla told the public that they could do that.

Lol, really? They either developed superman x-ray vision, or just tracked object occlusion with a common re-acquisition mitigation (so worthless when physical inertia carries a vehicle into an object collision.)

>So, it doesn't "grow exponentially",

The further the object... the more possible choices will need to be made in the guidance system. Note, guidance and navigation are related, but different problem domains. Roughly, the possible choices (and errors) if I recall grew by:

((m cars) * (k lanes ) * (r occlusions) * (s sign laws) * (1 + world_sate_delta_error(t)) ) ^ (n intersections + w way-points) = 1/hype_correction

...but that doesn't even cover the projected future risk(t). =3

I wonder if - at some point - AI might not do the "filling in" of information the brain does by generation (ie. filling in the gaps) or something.-

Maybe, but first a real "AI" will need to be invented.

I suspect a Amish horse buggy is more practical. =3

Failures in automotive programs are surely linear to time over a year with all seasons.

Failure rates are a different area, and most engineers have proven it is not linear. =3

https://en.wikipedia.org/wiki/Bathtub_curve

Human bicameral driven cars are unreliable too, but FSD is more reliable. And more reliable, with way more data than any competing technologies.

Using vision for driving is something that has worked for as long as cars have existed. Trying to push some "millimeter precision" solution with unproven feature set and prohibitive hardware accessibility is just asking for no real safety improvements and just more lives lost.

Cheers.

It is true most humans can not track more than 5 objects concurrently. However, lidar data sets tend to produce sparse unambiguous point cloud data that is a lot more feasible to process. For example, a picture/reflection of a road on a bus is not a viable option with 3D data.

Cheers =3

https://www.youtube.com/watch?v=vJG698U2Mvo

Adding some stat of x things tracked adding top of your stat of "millimeters" really shows the level you're arguing here. Real world doesn't care about exact stats on some simply enumerable attributes, but adaptability. You might see 10 bisons but will instantly react to a scorpion in middle of them in your vision.

In any case. Trying to argue against vision even if LIDAR hypothetically was better (it isn't) would just lead to more deaths, maybe at best shielding the rich driving in cities. FSD's stats don't lie :|

My point was LIDAR data removes the ambiguity of camera derived data, is not fooled by featureless area ambiguity, and it is computationally easier to handle sparse data.

Many high-end multi-beam lidar also embed things like basic Bicycle and Pedestrian object detection in the sensor front end. Things have improved significantly with sensors, but the risk is never 0.

The rate of death doesn't really override an expectation of product safety, and humans understanding other humans intent.

Have a wonderful day =3

Fragmentation of results removing instead of increasing ambiguity doesn't pass the smell check, neither isn't it not something exactly against the leading FSD engineers have commented on the issue while they were removing other non-uniform sensors, even disabling existing ones entirely in cars.

Trying to claim some entirely different stack in some third party LIDAR tower's own processing is somehow "beneficial" sounds like a project manager who thinks adding engineers equates to linearly faster progress.

Just no. On a slight tangent though, I recommend reading about Tesla's vertical integration. It's not something any other company has managed to get implemented so deep in automotive, which makes it quite incomparable in some aspects where others can't adapt even if they wanted to.

Admittedly, I only have limited out-of-date knowledge from integrating iBeo, hokuyo, and SiCK lidar models (way better models out now.) Yet in general, the SLAM algorithms tended to use optical Odometry, pose recovery, and point-cloud data. Then extrapolated the projected surface using camera data within a GPU accelerated OpenCV layer. It does require a proper global shutter machine vision camera (high fps / no smear), but does work fairly well under ideal conditions. The winners of the Darpa grand challenge documented the methods in detail, and several OpenCV books cite the work as a student project.

Let me know if you have trouble finding the projects. =3

gluing together third party stacks like there's no tomorrow doesn't make you better evaluating how a vision-only ML model with limited, yet uniquely hardware-accelerated, one-task actual real life, real-time ML model running right now would integrated with foreign data, especially as the actual relevant engineers have directly talked against such things.

It might make you headstrong in believing against something that'd be easier to see the core sensibility of if you weren't so invested in just a specific corner/angle though.

I've avoided working a work project involving LIDAR scanning before, even back then the hellishness of the hardware was a large factor. I wouldn't mind playing around with a Jetson Nano though.

>gluing together third party stacks like there's no tomorrow

That is why I don't really like ROS. lol =3

>doesn't make you better evaluating how a vision-only ML model

In general, the monocular SLAM algorithms rely on salient feature extraction, and several calibrated assumptions about the camera platform. How you interpret that output is another set of issues, as the power budget is going to take the hit.

For machine vision, I'd skip the proprietary Jetson Nano... and get a cheap gaming "parts" laptop with a broken LCD and several USB ports (RTX4090 or RTX4080 is a trophy.)

No one wants to fork over $30k for an outdoor lidar, but using only cameras is a fools errand. The best platforms I've seen commercially use camera + lidar + radar.

For student projects, one can get small radars and TOF sensors for under $20 off sparkfun (similar to the one in iPhone Pro 11/12/13). We live in the future... =3

> FSD is more reliable Citation needed. And not Tesla’s typical lying with shitty statistics data (fun fact: most car accidents do not happen in places where FSD is commonly used).

Also, vision-only systems work great… if they’re backed by strong intelligence.

Only cameras? I respectfully disagree.

https://www.youtube.com/watch?v=IQJL3htsDyQ

Your previous link was of that old BBC vision test, comparing being safe from danger to being able to see obviously irrelevant details of a video.

And now you link that debunked Mark Rober video that literally doesn't even have FSD turned on, while giving the most ridiculous free wins to LIDAR. Talk about writing the tests for the exact limits of a specific system. https://www.youtube.com/watch?v=QhX_fgekpk0

You're really running out of steam :D

Kind of sus, but it is consistent with other reported incidents including at least 1 fatality with FSD.

https://www.cnn.com/2025/01/07/business/nhtsa-tesla-smart-su...

Best of luck =3

"Consistent" comparing a video with a literal billboard of a road in the middle of a road to real life? With the entire FSD feature turned off in the video by the way :D ?

And one fatality with the most dangerous general form of transportation we have?

And an article about a fancy pants 5 miles a hour park retrieval feature bending a few posts as if it was relevant?

Dude I don't need luck, I could roll ten D12 ones in a row and win

One of the YT videos is probably incorrect, but other data under less demanding conditions already suggests FSD camera only options don't work as advertised within cities.

I'd rather not have drivers playing dice while driving. =3

some data suggests maybe this result in maybe some condition maybe

Tesla always seems to find its telemetry data when they get sued. Some of their customers are less then honorable. lol =3

Company doesn't spread all its private data all over the internet all the time

more at 5

So would you rather have the actual data fed to you by a different parent? Airplane goes brrrr ! https://insideevs.com/news/720730/tesla-autopilot-crash-data...

>fun fact: most car accidents do not happen in places where FSD is commonly used

How is that even supposed to in any way be relevant when talking exactly about cases where FSD and similar are used. Sigh.

>Also, vision-only systems work great… if they’re backed by strong intelligence.

Yes. And I do recognize that "Best, with custom in-house NN hw" might not still be "Strong" on all aspiratory statistics. But its already much above human capability, and regardless if you want to try to say the stats are 2x 3x, even 4x exaggerated, they'd still blow the alternative safety standard out of the water.

> Human bicameral driven cars are unreliable too, but FSD is more reliable.

You do not have the data necessary[0] to substantiate this claim.

[0] Accidents per mile controlled for at least vintage of car and driving conditions

I do have the data necessary[0] to substantiate this claim. And no, I'm not going down listening in the rabbit hole of people trying to poke pin prick holes in to a safety margin bigger than a Gigafactory.

[0] https://insideevs.com/news/720730/tesla-autopilot-crash-data...

1. Autopilot is not FSD

2. Autopilot, being a typical ASAD, is used exclusively on highways and in conditions where typical ASADs work reliably

Weird to brag about being unwilling to apply even first order criticality to a press release but you do you.

Thank you for reaffirming that in fact you do not have the data required to substantiate your claim.

As I said, I'm not going down the rabbit hole. You go ahead and drive right in to the fallen tree in the middle of the road as you so vehemently argue it's not there.

I think you should consider reasoning about reality in real terms instead of offloading to metaphors that, by design, don’t add any more information to the discussion than your own conclusions.

By the way, it’s a “binocular” system. “Bicameral” refers to a design for institutions like legislatures.

I didn't know Teslas come with a House of Lords

Probable, based on the cyber truck build... lol =3

Reply on news.ycombinator.com