Tesla has dropped a major bombshell saying it will make its own robot, called the Teslabot, using the same computing hardware it has developed for autonomous driving.
In an online event to outline the EV and energy storage maker’s artificial intelligence plans to use a “Dojo” neural network, Tesla’s AI team led by Andrej Karpathy laid out the path Tesla has taken to solve the huge computing requirements of autonomy.
And not just on wheels, it turns out.
“Arguably, Tesla is the world’s biggest robotics company because our cars are essentially semi-sentient robots on wheels,” CEO Elon Musk.
“Dojo should be operational next year,” he added.
About 5’8″ tall, able to carry around 20 kilos and deadlift almost 80 kilos, the Teslabot will be programmed to move no faster than 8 kilometres an hour, said Musk.
It would use all the same technology as the car, and “walk through the world without being explicity trained.”
“It should be able to go to the store and get the groceries. I think we can do that,” he said.
“If we don’t make it, someone else will. So we should, and make sure it’s safe,” he said.
We have almost all the pieces needed for humanoid robots, since we already make robots with wheels
— Elon Musk (@elonmusk) August 20, 2021
Arguing that the foundation of economy is labour, Musk mused about a universal basic income and a world in which “physical work will be a choice, if you want to you can do it but you won’t need to.”
Ideally, it would do boring, repetitive and dangerous jobs. “But not right now because the robot doesn’t work,” he added.
While that bombshell no doubt has a massive amount of unpacking to do, the core of Tesla’s AI Day was to attract the brightest brains to help the company solve the final problems needed for fully autonomous self-driving, which will enable it to achieve its plan of deploying fleets of robotaxis.
Taking the audience through the full gamut of its AI journey from single-camera and radar driven cars of less than a decade ago to radar-free, eight-camera Tesla Vision-capable cars of today, Tesla’s AI team explained the numerous problems it has been tasked with solving.
This has included combining vectors created from imagery captured by multiple cameras, using neural depth to allow, for example, chips “looking at” (processing) a few pixels in the distance and correlating it with the fact it sits at the vanishing point of the highway to determine it is in fact a car.
The team talked about they have developed vehicles that can recognise and process the curbs and lines on a road, integrating vision from the vehicle to create birds-eye view predictions of what the road in its surrounds looks like.
They discussed moving away from labelling of objects in a 2D environment by a third-party, to 4D predictive labelling in time and space.
“Honestly the quality was not amazing,” said Karpathy.
Instead, the Tesla AI team developed an AI chip described in 2018 by Musk as “super kick-ass”, with the lowest latency possible and extremely high bandwidth deployed in a supercomputer called the Exapod.
Made of 500,000 tile configuration with nine petaflops per tile, processing 36 terabytes per second, this is just the tip of the iceberg for Dojo.
With Project Dojo’s Exapod computer, Venkataramanan says that Tesla will be able to implement automatic labelling of objects, doing away with manual labelling by humans.
Tesla also plans to increase its rate of simulations to train its AI software to recognise and react safely to situations it would not normally come across “in the wild,” or is difficult to label.
To date, the company has created 371 million simulated images and 480 million cuboids to use in its simulations.
The Dojo computer will literally be a “label factory,” says Venkataramanan.
“It will be the fastest AI training computer, with 4x the performance at the same cost, 1.3x energy saving performance and 5x smaller footprint. This will be Dojo computer,” he says.
“But we are not done. We are assembling our first cabinets pretty soon and have a whole next generation plan we are already thinking about,” he said.
Bridie Schmidt is associate editor for The Driven, sister site of Renew Economy. She has been writing about electric vehicles since 2018, and has a keen interest in the role that zero-emissions transport has to play in sustainability. She has participated in podcasts such as Download This Show with Marc Fennell and Shirtloads of Science with Karl Kruszelnicki and is co-organiser of the Northern Rivers Electric Vehicle Forum. Bridie also owns a Tesla Model Y and has it available for hire on evee.com.au.