[dgindia]

In the yesteryears, we have witnessed fierce battle in video industry on the underlying technology of PAN vs NTSC. I realised when a camcorder bought in US (having PAL technology) was not functioning when connected to Indian TV (having NTSC technology). We had to avail services of experts who would burn CD is compatible technology for viewing at home.

Something similar is brewing in the automobile industry - the rival technologies used for self-driven card are LIDAR and FSD.

LIDAR (Light Detection and Ranging) is usage of laser blight to sense the objects in surrounding and collect the information to create a 3D digital image. Many car manufacturers like Mercedes, Nissan, BMW, VW, Volvo are using this technology in their pursuit of self-propelled cars.

FSD (Light Detection and Ranging) is another technology using cameras and computer vision to generate AI driven surrounding's digital clone. Tesla has pioneered this technology and is a key user in their products.

Critics say LIDAR supports multi driving conditions when it is dark. FSD is more cost effective and mimics human driving behaviour.

https://medium.com/technology-hits/e...r-f8b1ef0d024c

It will be interesting to witness the future of driving is a high-stakes battle between lasers and cameras.

Given both options, which car will you choose to relax while being auto-driven.

[MotoBlip]

Automated driving success hinges not just on sensor quantity or tech choice, but on aligning hardware with the operational design domain (ODD) requirements. Achieving advanced automated driving capabilities isn't simply about incorporating numerous sensors or cutting-edge technology—it's about ensuring the hardware precisely aligns with the vehicle's requirements for safe and effective operation in real-world conditions. Consider Tesla's FSD Beta system, designed to eventually achieve full autonomy (Level 4/5), yet currently labeled as Level 2, potentially to navigate strict regulatory frameworks.

In China, companies like NIO, Xpeng, and Xiaomi are integrating LIDAR into their systems, though the current capabilities and potential improvements through software updates remain somewhat ambiguous. Levels 2/3 of automated driving necessitate human oversight, requiring a person to be prepared to assume control if the system encounters difficulties. This requirement persists regardless of the vehicle's sensor array.

It's worth noting that simply adding sensors beyond the initial design won't inherently enhance the vehicle's capabilities beyond its original specifications. The complexity of navigating urban environments with ADAS already presents significant challenges, irrespective of the diversity or quantity of sensors employed.

From my experience working with LIDAR sensors, I've observed certain limitations. Their performance can degrade in adverse weather conditions, with time-of-flight(ToF) LIDAR sensors particularly affected compared to frequency-modulated continuous wave(FMCW) sensors, though the latter come at a significantly higher cost. Another consideration is the choice between short-range or long-range LIDAR, or both, which adds to the system's overall cost and complexity. It's definitely not as simple as some people make it out to be.

[Rehaan]

Originally Posted by dgindia

Critics say LIDAR supports multi driving conditions when it is dark. FSD is more cost effective and mimics human driving behaviour.

Haven't researched this in too much detail, but I've also heard that LIDAR does not work well in the rain, and also it could become an issue when there are dozens of LIDAR devices all running simultaneously in a small area.


For me, the ultimate choice is always "sensor fusion". The combination of inputs from several different types of sensors (each of which has their strong and weak points).

Think of humans -- not only are we using our eyes to drive, but also taking inputs from our ears and hands/butt, sense of balance, and even sometimes our noses.

However, this approach is difficult because:
1) To code it (ie. which sensor to give priority to when there's a conflict) gets more and more complicated the more sensors you add
2) It increases hardware cost as well as R&D cost

[NomadSK]

FSD/Autonomous driving is ages away. Not happening soon.

All sensors need to be calibrated be it LIDAR, camera based, IR or ultrasonics (RADAR/SONAR), radiography, if you want 100% accuracy. They generally work near perfectly in controlled environments, outside of that envelope its very difficult to make them work according to your needs. Each technology has their own set of limitations.

We use all these sensors for various inspections/investigation techniques. They don't give the exact results but have to be inferred by SME's, with approximations. FSD and auto drive i.e level - 5 of ADAS is a farce at the moment and is probably at least a decade away, if not more. We have still not reached completely at level 3. You can reach very close, but not there. Theoretically, the only way it can be achieved safely when 100% of the cars will have similar features, else there are too many variables in the real world to come close to fully autonomous driving.

I'm of the opinion if sensors based autonomous travel will start, the first industry to incorporate, will be aviation industry, they are super close to full automation and having dedicated landing/take-off facilities. Else there's a reason train runs on rails.

[sumitj71]

I remember once Tesla has said that they went with camera based self driving system because our road environment is made for us, we heavily use visual information for driving. All road signs and signals are designed for the same purpose. LIDAR can not detect the symbols in any road sign or can not detect signals. We process visual information with our intelligence developed over time to navigate driving invironment. Hence visual it IR camera based self driving is preferred by Tesla.
Second advantage of camera based self driving is it is a passive system, it does not emit anything which can interfere with any other similar system.
I think camera based self driving is possible with AI.

[RobotMan]

Having worked with both cameras and LIDAR, I think the challenge is going to be in compute power.

Just like any problem can be solved with money, similarly, any computational challenge can be solved by throwing enough compute at it. Sadly, it takes more compute power than is currently available on cars.

For current Level 4 self driving cars (specifically Waymo tech stack) there is the LIDAR, the RADAR and the Cameras. Each one cover's up the weakness of the other for ex: LIDAR is crappy in rain, Cameras have bad depth perception and can be fooled like by images, Radar is less accurate but has the maximum range. All three systems together generate huge quantities of high quality data in real time. Waymo also uses HD maps which are used by the vehicle to localize itself in the surrounding. Its up to the onboard computers (CPU + GPU) to clean up this data, fuse it, process it and then act on it. Doing all this in real time is possible. Industries already do it in closed, controlled environments.

Performing all these operations in real time, in unpredictable surroundings, with limited compute is really tough. That's one of the biggest reasons Waymo is so slow with its roll out. They first generate HD maps, then train their software stack on a per city basis, test, test, test some more, release in beta and then release to public.

This approach takes time, all these sensors are extremely expensive and Waymo seems to be following the 'technology will catch up' methodology. They are hoping sensors will become cheap in the future. They are hoping compute will become cheap in the future. A Waymo car from San Francisco cannot operate directly in Arizona. It needs a software update with Arizona data/models. The cars are geo-fenced. They are not profitable right now.

Talking about vision only systems (Tesla, Comma.AI) these only use cameras (in Tesla's case, cheap cameras). With enough processing, vision can be used during night time and in bad weather. But training the AI model for driving just using vision needs a lot of data. It becomes extremely compute intensive. That's why Tesla needs it super computer. Their onboard Hardware 3.0 computer while quite capable, is no where near as capable to handle every scenario that the super computer based AI model has been trained on. As a result, Tesla focusses only on California (and now Texas) streets. Their FSD systems performs way better in these areas than in other cities/localities in terms of interventions and disengagements. No matter how much Elon sings praises of his 'generalized AI model' which can handle any and every scenario, it can't. Not with the current hardware.

A small note about AI/ML models: every company (Waymo, Tesla, Zoox etc) has realized AI is the way to make the self driving car a reality. For this discussion, I am using Tesla's software as an example as they have the most amount of training data (6 billion+ miles).
Tesla teaches its models what is good behavior and what is bad behavior. If they train their models on all the data they have available, the models will have trillions of parameters, will be terabytes in size and will be completely unmanageable. It would be next to impossible to run these models on the car's computer and even sending it OTA will be a nightmare
That's why Tesla is extremely selective about their training data. Sometimes they miss a few scenarios and people see regressions/regressive behavior.

In summary, barring cost, software has proven that both approaches can work. The math behind both approaches checks out. The race currently is to prove who can get there faster and profitably. Without significant improvement in compute capacity and cost reduction in sensors, neither of the approaches are going to work.

[BullettuPaandi]

I would bet on FSD, but that's only considering the disproportionate investment in R&D on tech that'd make FSD better over time, than LIDAR. If we consider just the underlying technologies and their theoretical potential, it really looks like a cat on the wall to me.

But realistically, I'd go with neither to the question of which would I choose to be driven on. I'd just go with public transport or a cab/ chauffeur depending on things like if I'm alone or with family, luggage, budget, etc. - better peace of mind, virtually infinite ODD and reasonable costs. I'm just glad that there are entities that continue to invest in this novelty of auto-driven cars, only because any improvement there also means betterment for the rest- safety features get trickled down to regular automotives, better infrastructure (which I believe is essential to improve upon the ODD), etc. I'm just waiting for that day, when the governments are forced to improve infrastructure, to keep selling cars, given that this turnover seems to be a part of the economic path we've chosen.

[ph03n!x]

Thread's title is incorrect/ misleading - LiDAR is a laser-based method for mapping surfaces and objects, and requires LiDAR sensors. FSD is Tesla's way of calling autonomous driving that matches SAE L5 without using LiDAR sensors, but using regular, visible light cameras.

I think what the OP is debating is dedicated, LiDAR sensors augmented with camera imaging vs. pure camera imaging as implemented by Tesla FSD.

I'm technically challenged in this aspect, but looks like Tesla's neural network based FSD, which does not use LiDAR, works relatively well, and at scale compared to say, Waymo or Cruise. That could be because of the discipline or type of traffic where it's implemented at scale.

For conditions like India, remains to be seen if FSD kind of an implementation is successful, or will LiDAR (or similar) sensor based augmentation makes it better.

[DicKy]

Originally Posted by dgindia

In the yesteryears, we have witnessed fierce battle in video industry on the underlying technology of PAN vs NTSC. I realised when a camcorder bought in US (having PAL technology) was not functioning when connected to Indian TV (having NTSC technology). .

Sorry for going way offtopic. Wasn't the Americas that used NTSC and we used PAL?

Cause I remember the products we used in Saudi Arabia had options to chose between PAL, NTSC and SECAM and I remember we using PAL both in Saudi Arabia and India, just that we had to take care of the 110v, 220v difference and bring adopters along.

A more apt example would be VHS vs Betamax wars as it was market driven rather than government regulation dictated.

[Jeroen]

Originally Posted by DicKy

A more apt example would be VHS vs Betamax wars as it was market driven rather than government regulation dictated.

And the world ended up with VHS, whereas Betamax was the technical superior technique. I am all for free markets, but it doesn’t guarantee the best solution. It will guarantee certain individuals and companies to make tonnes of money.

Jeroen

[DicKy]

Originally Posted by Jeroen

And the world ended up with VHS, whereas Betamax was the technical superior technique. I am all for free markets, but it doesn’t guarantee the best solution. It will guarantee certain individuals and companies to make tonnes of money.

Jeroen

Of course yes. Betamax was the superior one, which was why SONY kept believing in it.

I meant it as a reply to OP for an example of LIDAR vs FSD, the Betamax vs VHS would be more suitable, unlike PAL vs NSTC that was more government dicatated and limited by political boundaries.

[Quatro]

In the Indian environment, I believe Honda is the only one where its ADAS is based on Camera, whereas everyone else seems to have both - Lidar and Camera. For an important issue like this and that can be fatal also, I will prefer having a combination of both options rather than relying on only one of the two.

[dgindia]

Wasn't the Americas that used NTSC and we used PAL?

I stand corrected. Thank you.