A supercomputer trains Tesla to drive with just cameras
At the Conference on Computer Vision and Pattern Recognition, Tesla presented his new supercomputer which will allow the autonomous driving only through video cameras. According to the head of Artificial Intelligence Andrej Karpathythis technology will improve driving while giving up the LIDAR and radar sensors.
Tesla’s new supercomputer
Since 2019, Elon Musk has been promoting Tesla’s future neural network, a supercomputer called Dojo. Treating cars like so many electric Karate Kids, Dojo will train driving using only cameras. The supercomputer presented yesterday is not quite as advanced but rather acts as a predecessor and test for this technology. For a car to respond to the environment as a man would, it needs to collect and processor one huge amount of data.
How many? The new Tesla supercomputer has 10 petabytes of NVME memory and can process up to 1.6 terabytes per second. With 1.8 EFLOPS of power ranks fifth among the most powerful computers on Earth, although the company has not tested the benchmarks necessary to officially enter the ranking.
The advantage of the video camera
Many companies offering self-driving vehicles use highly detailed maps and LIDAR sensors. But Musk’s engineers prefer to focus on cameras for one simple reason: latency. Visual feedback is faster and more immediate, potentially much better than human reflexes, because he can’t get distracted. But if our brains have had millions of years of evolution to analyze the environment around us so efficiently, supercomputers are speeding up the analysis by a lot.
Starting in May, Tesla Model Y and Model 3 built in North America they no longer have LIDAR sensors but they rely solely on video cameras. And they’ve already added a safety feature called “Pedal Wrong Application Mitigation”, which identifies pedestrians in front of the car and brakes even if by mistake the driver presses the accelerator in panic. Tesla assures us that we will see more and more of these kinds of functions, as artificial intelligence is refined.
The car maker’s new supercomputer has amassed a million videos taken by eight cameras around the vehicles, lasting around 10 seconds. This made it possible to tag six billion objects in depth, speed and acceleration. In total, they are 1.5 petabytes of memory. And this is just the beginning: the company aims to improve these characteristics in the near future. Why, as he explains Karpathy“Still struggles in complicated environments like San Francisco.”
But Tesla has chosen the camera-only strategy and I start the road to get there. Now he has to start running.