Tesla AI chief explains why self-driving cars don’t need lidar

9 min read
Tesla dashboard

What is the technology stack you need to create fully autonomous vehicles? Companies and researchers are divided on the answer to that question. Approaches to autonomous driving range from just cameras and computer vision to a combination of computer vision and advanced sensors.

Tesla has been a vocal champion for the pure vision-based approach to autonomous driving, and in this year’s Conference on Computer Vision and Pattern Recognition (CVPR), its chief AI scientist Andrej Karpathy explained why.

Speaking at CVPR 2021 Workshop on Autonomous Driving, Karpathy, who has been leading Tesla’s self-driving efforts in the past years, detailed how the company is developing deep learning systems that only need video input to make sense of the car’s surroundings. He also explained why Tesla is in the best position to make vision-based self-driving cars a reality.

A general computer vision system

Deep neural networks are one of the main components of the self-driving technology stack. Neural networks analyze on-car camera feeds for roads, signs, cars, obstacles, and people.

But deep learning can also make mistakes in detecting objects in images. This is why most self-driving car companies, including Alphabet subsidiary Waymo, use lidars, a device that creates 3D maps of the car’s surrounding by emitting laser beams in all directions. Lidars provided added information that can fill the gaps of the neural networks.

However, adding lidars to the self-driving stack comes with its own complications. “You have to pre-map the environment with the lidar, and then you have to create a high-definition map, and you have to insert all the lanes and how they connect and all the traffic lights,” Karpathy said. “And at test time, you are simply localizing to that map to drive around.”

It is extremely difficult to create a precise mapping of every location the self-driving car will be traveling. “It’s unscalable to collect, build, and maintain these high-definition lidar maps,” Karpathy said. “It would be extremely difficult to keep this infrastructure up to date.”

Tesla does not use lidars and high-definition maps in its self-driving stack. “Everything that happens, happens for the first time, in the car, based on the videos from the eight cameras that surround the car,” Karpathy said.

The self-driving technology must figure out where the lanes are, where the traffic lights are, what is their status, and which ones are relevant to the vehicle. And it must do all of this without having any predefined information about the roads it is navigating.

Karpathy acknowledged that vision-based autonomous driving is technically more difficult because it requires neural networks that function incredibly well based on the video feeds only. “But once you actually get it to work, it’s a general vision system, and can principally be deployed anywhere on earth,” he said.

With the general vision system, you will no longer need any complementary gear on your car. And Tesla is already moving in this direction, Karpathy says. Previously, the company’s cars used a combination of radar and cameras for self-driving. But it has recently started shipping cars without radars.

“We deleted the radar and are driving on vision alone in these cars,” Karpathy said, adding that the reason is that Tesla’s deep learning system has reached the point where it is a hundred times better than the radar, and now the radar is starting to hold things back and is “starting to contribute noise.”

Supervised learning

Tesla object detection

The main argument against the pure computer vision approach is that there is uncertainty on whether neural networks can do range-finding and depth estimation without help from lidar depth maps.

“Obviously humans drive around with vision, so our neural net is able to process visual input to understand the depth and velocity of objects around us,” Karpathy said. “But the big question is can the synthetic neural networks do the same. And I think the answer to us internally, in the last few months that we’ve worked on this, is an unequivocal yes.”

Tesla’s engineers wanted to create a deep learning system that could perform object detection along with depth, velocity, and acceleration. They decided to treat the challenge as a supervised learning problem, in which a neural network learns to detect objects and their associated properties after training on annotated data.

To train their deep learning architecture, the Tesla team needed a massive dataset of millions of videos, carefully annotated with the objects they contain and their properties. Creating datasets for self-driving cars is especially tricky, and the engineers must make sure to include a diverse set of road settings and edge cases that don’t happen very often.

“When you have a large, clean, diverse datasets, and you train a large neural network on it, what I’ve seen in practice is… success is guaranteed,” Karpathy said.

Auto-labeled dataset

Tesla data engineering cycle

With millions of camera-equipped cars sold across the world, Tesla is in a great position to collect the data required to train the car vision deep learning model. The Tesla self-driving team accumulated 1.5 petabytes of data consisting of one million 10-second videos and 6 billion objects annotated with bounding boxes, depth, and velocity.

But labeling such a dataset is a great challenge. One approach is to have it annotated manually through data-labeling companies or online platforms such as Amazon Turk. But this would require a massive manual effort, could cost a fortune, and become a very slow process.

Instead, the Tesla team used an auto-labeling technique that involves a combination of neural networks, radar data, and human reviews. Since the dataset is being annotated offline, the neural networks can run the videos back in forth, compare their predictions with the ground truth, and adjust their parameters. This contrasts with test-time inference, where everything happens in real-time and the deep learning models can’t make recourse.

Offline labeling also enabled the engineers to apply very powerful and compute-intensive object detection networks that can’t be deployed on cars and used in real-time, low-latency applications. And they used radar sensor data to further verify the neural network’s inferences. All of this improved the precision of the labeling network.

“If you’re offline, you have the benefit of hindsight, so you can do a much better job of calmly fusing [different sensor data],” Karpathy said. “And in addition, you can involve humans, and they can do cleaning, verification, editing, and so on.”

According to videos Karpathy showed at CVPR, the object detection network remains consistent through debris, dust, and snow clouds.

tesla object tracking auto-labeling
Tesla’s neural networks can consistently detect objects in various visibility conditions.

Karpathy did not say how much human effort was required to make the final corrections to the auto-labeling system. But human cognition played a key role in steering the auto-labeling system in the right direction.

While developing the dataset, the Tesla team found more than 200 triggers that indicated the object detection needed adjustments. These included problems such as inconsistency between detection results in different cameras or between the camera and the radar. They also identified scenarios that might need special care such as tunnel entry and exit and cars with objects on top.

It took four months to develop and master all these triggers. As the labeling network became better, it was deployed in “shadow mode,” which means it is installed in consumer vehicles and run silently without issuing commands to the car. The network’s output is compared to that of the legacy network, the radar, and the driver’s behavior.

The Tesla team went through seven iterations of data engineering. They started with an initial dataset on which they trained their neural network. They then deployed the deep learning in shadow mode on real cars and used the triggers to detect inconsistencies, errors, and special scenarios. The errors were then revised, corrected, and if necessary, new data was added to the dataset.

“We spin this loop over and over again until the network becomes incredibly good,” Karpathy said.

So, the architecture can better be described as a semi-auto labeling system with an ingenious division of labor, in which the neural networks do the repetitive work and humans take care of the high-level cognitive issues and corner cases.

Interestingly, when one of the attendees asked Karpathy whether the generation of the triggers could be automated, he said, “[Automating the trigger] is a very tricky scenario, because you can have general triggers, but they will not correctly represent the error modes. It would be very hard to, for example, automatically have a trigger that triggers for entering and exiting tunnels. That’s something semantic that you as a person have to intuit [emphasis mine] that this is a challenge… It’s not clear how that would work.”

Hierarchical deep learning architecture

Tesla neural network self-driving car

Tesla’s self-driving team needed a very efficient and well-designed neural network to make the most out of the high-quality dataset they had gathered.

The company created a hierarchical deep learning architecture composed of different neural networks that process information and feed their output to the next set of networks.

The deep learning model uses convolutional neural networks to extract features from the videos of eight cameras installed around the car and fuses them together using transformer networks. It then fuses them across time, which is important for tasks such as trajectory-prediction and to smooth out inference inconsistencies.

The spatial and temporal features are then fed into a branching structure of neural networks that Karpathy described as heads, trunks, and terminals.

“The reason you want this branching structure is because there’s a huge amount of outputs that you’re interested in, and you can’t afford to have a single neural network for every one of the outputs,” Karpathy said.

The hierarchical structure makes it possible to reuse components for different tasks and enable feature-sharing between the different inference pathways.

Another benefit of the modular architecture of the network is the possibility of distributed development. Tesla is currently employing a large team of machine learning engineers working on the self-driving neural network. Each of them works on a small component of the network and they plug in their results into the larger network.

“We have a team of roughly 20 people who are training neural networks full time. They’re all cooperating on a single neural network,” Karpathy said.

Vertical integration

Tesla AI computers

In his presentation at CVPR, Karpathy shared some details about the supercomputer Tesla is using to train and finetune its deep learning models.

The compute cluster is composed of 80 nodes, each containing eight Nvidia A100 GPUs with 80 gigabytes of video memory, amounting to 5,760 GPUs and more than 450 terabytes of VRAM. The supercomputer also has 10 petabytes of NVME superfast storage and 640 tbps networking capacity to connect all the nodes and allow efficient distributed training of the neural networks.

Tesla also owns and builds the AI chips installed inside its cars. “These chips are specifically designed for the neural networks we want to run for [full self-driving] applications,” Karpathy said.

Tesla’s big advantage is its vertical integration. Tesla owns the entire self-driving car stack. It manufactures the car and the hardware for self-driving capabilities. It is in a unique position to collect a wide variety of telemetry and video data from the millions of cars it has sold. It also creates and trains its neural networks on its proprietary datasets, its special in-house compute clusters, and validates and finetunes the networks through shadow testing on its cars. And, of course, it has a very talented team of machine learning engineers, researchers, and hardware designers to put all the pieces together.

“You get to co-design and engineer at all the layers of that stack,” Karpathy said. “There’s no third party that is holding you back. You’re fully in charge of your own destiny, which I think is incredible.”

This vertical integration and repeating cycle of creating data, tuning machine learning models, and deploying them on many cars puts Tesla in a unique position to implement vision-only self-driving car capabilities. In his presentation, Karpathy showed several examples where the new neural network alone outmatched the legacy ML model that worked in combination with radar information.

And if the system continues to improve, as Karpathy says, Tesla might be on the track of making lidars obsolete. And I don’t see any other company being able to reproduce Tesla’s approach.

Open issues

But the question remains as to whether deep learning in its current state will be enough to overcome all the challenges of self-driving. Surely, object detection and velocity and range estimation play a big part in driving. But human vision also performs many other complex functions, which scientists call the “dark matter” of vision. Those are all important components in the conscious and subconscious analysis of visual input and navigation of different environments.

Deep learning models also struggle with making causal inference, which can be a huge barrier when the models face new situations they haven’t seen before. So, while Tesla has managed to create a very huge and diverse dataset, open roads are also very complex environments where new and unpredicted things can happen all the time.

The AI community is divided over whether you need to explicitly integrate causality and reasoning into deep neural networks or if you can overcome the causality barrier through “direct fit,” where a large and well-distributed dataset will be enough to reach general-purpose deep learning. Tesla’s vision-based self-driving team seems to favor the latter (though given their full control over the stack, they could always try new neural network architectures in the future). It will be interesting to how the technology fares against the test of time.

10 COMMENTS

  1. Most of what they’re doing is fairly obvious. Even going back some 20 years while I was studying AI, I wrote down many similar ideas on the subject. Having studied Physics prior to AI, I suppose I had an advantage over most CS students (especially in those days) on how to deal with things like space, time, momentum and causality.

    Nowadays many of these ideas have been implemented, including by Tesla, but they’re making the same mistakes that were made decades ago in AI that caused it to fall out of favor even before I decided to enter the field. Companies then and now try to fly before they can crawl, and high-level intelligence simply doesn’t work that way. Even a human isn’t capable of driving a car without the ability to do many other things their system lacks. Yet they insist on the comparison with humans when it comes to a vision-only model. That would be fine if they could replicate all those other things humans need to learn before driving, but that’s not the case.

    I wouldn’t dream of trying to implement these models on unconstrained heavy machinery such as a motor vehicle without having done years of validation with less dangerous, smaller scale vehicles/robots in a production environment involving interaction with humans and human-controlled devices.

    Driving necessarily involves communicating (primarily non-verbally) with and implicitly understanding intent of other humans on the road, as pedestrians, operators and officials. It’s not just a matter of blindly following the rules of the road. The so-called “edge cases” cannot be reduced to a finite number when considered as such. Situations like those are an integral part of our human experience and also why we do not allow people to drive until reaching a certain age, though they may otherwise be perfectly capable of the technical aspects of driving.

    The most irresponsible part is that I’m quite sure they know better but do it anyway despite the cost to others. They demonstrate this via their disclaimers and how they use their paying customers (and others on the road) as guinea pigs while taking no responsibility for whatever happens as a result.

    They should also know that the statistics they cite to sway public opinion are garbage. First off, one can’t compare the safety record of their system with that of human operators because they constrain when the system is autonomous and require humans to take over whenever it fails. If you could isolate the easiest driving situations for humans and determine a safety record from that alone, those statistics would be quite different.

    Furthermore, safety must be considered in the context of productivity. In other words, being safe is meaningless if to do so would cost people an unacceptable amount of time to reach their destinations. That not only includes their customers, but those in other vehicles. Their system, if widely implemented with full autonomy, would almost certainly create traffic nightmares as misunderstandings between their vehicles and human operators lead to deadlock/gridlock in many cases. The only workaround would be to allow their vehicles to take greater risk to avoid that, but then the safety record would suffer greatly. They currently benefit from the low density of deployment, but that would change dramatically as the numbers increase.

    I suppose I can understand why so many people think Tesla will be successful in mass producing fully autonomous vehicles and dominate the industry. It’s likely because they don’t understand the first thing about AI or just how much goes into their own thought processes that they take for granted. They are easily manipulated by an immoral company run by an even worse excuse for a human being who happens to understand things only just enough to fool them into trusting him, when in reality they should be distancing themselves as quickly as possible.

    • You nailed it. I work in the tech industry and it is mind numbing how ignorant most of the industry is on things like this. Tesla thinks they just need massive amounts of data to train models capable of completely autonomous driving. Sure, it can get better than they have today with enough data, ground truthing, and model tuning, but it’s just pattern matching and basic heuristic rules at the end of the day. The moment you hit a scenario where enough inputs aren’t correlating to what’s been trained on, it goes to absolute shit. People argue “but it’s highly unlikely to happen”, yet it isn’t. Humans are able to rationalize far better on new visual inputs than a trained visual model and that’s because the human brain isn’t just pattern matching. I’m not sure why people call tech like Tesla’s autonomous driving solution “artificial intelligence” because intelligence is faaaaar more than pattern matching and heuristic rules. We are at least a decade away from actually having what a reasonable person would call artificial “intelligence”. I have to laugh when people talk about Skynet from the Terminator series being possible today because it shows how completely uninformed they are; we are still sooooo far from that.

  2. Karpathy’s opening argument leans on the notion that using lidars means pre-mapping the environments with said lidars, but that notion is completely false. There are lidar solutions right now that already have the ability to detect the road, lanes and stationary/non-stationary obstacles as well as classify said obstacles into categories. There is no need to pre-map the driving area in order to use lidar.

    Tesla as a whole also seem to be regarding camera-based vision and lidar-based vision as mutually exclusive solutions, and so they focus on explaining why cameras are better. But they are not mutually exclusive. These two technologies are complementary, each excelling where the other fails. Cameras are cheaper, more widely available, have better resolution and it is much easier to train neural networks on camera data, but they will always be susceptible to lighting condition changes and bad weather, whereas lidar is mostly unaffected by these factors. Using both technologies is the only way to properly have sensor redundancy, which is absolutely necessary when talking about autonomous vehicles responsible for human lives. Having multiple cameras doesn’t fulfill that condition when they can all fail at once due to the same cause (like exiting a tunnel into sunlight).

    I understand Tesla’s (and Musk’s) reason for their stance on this. They sold tens of thousands of cars equipped only with cameras (and radars) and promised customers and investors these cars will get L3 autonomous driving through software updates. They’ve committed to lidar-less L3, and admitting their mistake would make a lot of people angry. But you can’t help but hear their ulterior motive of not hurting their stock value in every statement or article about them.

  3. Anyone who has driven a Tesla (like my Model 3 Performance) in wet UK weather will know that camera-only autonomy is all but impossible. “Multiple cameras require cleaning” is a very popular dash message when driving on a wet motorway.

  4. That’s nice and all, but those paying attention will remember that when this changeover happened the new system wasn’t even ready and cars lost some of their capability. It’s ridiculous to talk about how they don’t need it when the transition resulted in them losing capability.

    They keep telling us how superior the AI system is while they slip farther and farther back from their timeline. Remember, we were supposed to have streets filled with self-driving Tesla robo taxis by now. Instead what we have is a so-called full self-driving system that they rush to explain is anything but to regulators.

    The first level 3 driver assist came from Honda. Tesla gets all the hype, but they are behind. The current state of their self-driving system is a massively complicated and unreliable toy. I am underwhelmed by attempts to hype how great their AI is and how it’s going to fix everything in the future because I’ve been hearing this from them for years, but it never happens.

  5. “open roads are also very complex environments where new and unpredicted things can happen all the time.”
    That’s when accidents happen, being them autonomous or human.
    Only with computers you can make them happen once for everybody. Humans are much more stubborn and autonomous.

  6. So they’re saying that humans don’t have lidar, and therefore, the cars don’t need it either. But humans do form memories of their surroundings and then use those memories to navigate. It’s easier to drive on familiar streets than on unfamiliar ones. Humans CAN drive on unfamiliar streets, but it’s more stressful.

Leave a Reply to Tundrok Cancel reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.