Why Tesla is taking a different approach to self-driving cars
Electric carmaker has faith that AI will beat Lidar and maps in steering vehicles
Elon Musk does not mince words when rejecting the technologies that other companies are relying on to control their driverless cars.
“The two main crutches that should not be used — and, in retrospect, will be obviously false and foolish — are Lidar and HD maps. Mark my words,” the Tesla chief executive said recently.
Going against conventional wisdom and jettisoning things most of his rivals see as essential sounds risky. But Mr Musk has never been one to follow the herd, or to under-promise.
The electric car maker says its new cars already have sufficient sensors and computers to drive themselves, and that it will send out an over-the-air software update before the end of the year to complete the picture (although it might take some time before insurance companies and regulators are willing to allow the cars to be used in fully autonomous mode.)
The “crutches” that Mr Musk complained about involve two of the most common ways for autonomous vehicles to understand the world around them.
Lidar sensors, which use lasers to send out pulses of light and measure the time it takes for a reflection to come back, are currently one of the best ways to measure the shape and distance of other objects. But they are expensive, with today’s mechanical models costing several thousand dollars.
Reducing their workings to a silicon chip might help. The world has become used to the cost of silicon components like this falling rapidly over time. But big cost reductions depend on producing in large volume.
Sharp reductions in the cost of smartphone components, for instance, reflect the scale of the market, with 1.5bn handsets sold in 2018. By contrast, only around 82m cars were sold around the world last year, and it is likely to be years before a significant fraction of new vehicles are equipped to be driverless.
High-definition maps, meanwhile, are used to help driverless cars understand their surroundings, reducing the amount of raw data they need to collect and process. This makes it necessary to “geofence” them, only allowing them to travel in areas that have been very precisely mapped.
The problem with this, according to Mr Musk, is that when the real world changes in a way that is not reflected in the map, it can cause the system to fail, and if you cannot count on the map to be 100 per cent accurate, it loses its value.
Instead of techniques like these, Tesla’s autonomous driving technology relies almost entirely on teaching its cars to “see” using an array of cameras. As a back-up, its cars also use a forward-facing radar, along with a dozen ultrasonic sensors around the vehicle to help detect objects that are close by.
Improvements in computer image recognition have already been among the biggest recent advances in artificial intelligence. As with much about AI, however, tasks that seem deceptively easy for humans can floor even the best computers.
Tesla has developed a combination of hardware and software to handle the task. Two weeks ago it revealed a computer chip it designed in-house to process the massive amounts of image data needed to enable its cars to interpret their surroundings. Even Nvidia, whose chips are widely used in the AI systems in other driverless cars, credited the company with “raising the bar” for the industry.
To make sense of all the data it is processing, Tesla relies on an AI technique called deep learning. This employs artificial neural networks, systems that were originally modelled on the visual cortex of animals.
This is where things get hard. Neural networks need large amounts of data to train on: only when they have been fed many different examples of the same thing, each painstakingly labelled by humans, can they learn enough to identity the same object in the real world.
For this, Tesla is counting on having a better real-world data set than any other company. It can draw on images captured by 400,000 or so cars that are already on the road (it hopes to get to 1m by around the middle of next year). With a piece of software it calls “shadow mode” operating in the background, it can track the behaviour of cars being driven by humans, using this to feed its learning systems.
But even this may not be enough. Neural networks are notoriously “brittle”: they can fail unexpectedly when, say, the image of an object does not match any of the variations they have been shown before. And there is always a risk that they will encounter a real-world situation they have not been trained for.
Despite this, Tesla is taking a defiantly purist approach. Feed the network enough real-world data about all the situations it could possibly encounter, it says, and the machine can match and surpass human drivers.
Take driving in snow, a notoriously difficult skill for driverless cars to master. Humans are surprisingly good at anticipating where lane markings are on snowy roads, says Andrej Karpathy, Tesla’s head of computer vision. Feed enough human-labelled images of snowy roads into an AI system, and the computer will eventually be able to interpret a similar scene, he says.
Others in the AI world question whether it is as straightforward as this makes it sound. According to one expert who recently had a close-up view of Tesla’s AI, the technology is still far short of being able to do what Mr Musk has promised for the end of this year, although this person adds that Tesla could well reach its goal in the next two to three years.
To judge from frequent delays with other Tesla products, Mr Musk often views deadlines as moveable commitments. But if he can reach full autonomy before others using a different technology, any delays may not matter.