Don't Laugh Off the Tesla Bot. Elon Musk's Optimus Is the Real Deal
Commentary: I saw the humanoid robot firsthand. Sure, Tesla's prototypes can't do flips, but the company has the right engineering approach for progress.
When watching Tesla's humanoid Optimus robot debut at AI Day 2022, it's entirely fair to be skeptical about the company's ultra-ambitious plans. The prototype robot's shuffling gait paled to the exciting parkour and flips of Boston Dynamics' Atlas, and Tesla has missed many deadlines bringing its full self-driving technology to its cars.
But I was at the AI Day event, and I'm here to advise you not to dismiss Tesla's sci-fi inspired vision.
The company revealed abundant evidence that it's thinking deeply about a bipedal robot and the artificial intelligence technology needed to make it useful. And much of Tesla's track record of genuine achievements in electric vehicle engineering and manufacturing apply to Optimus.
Non-Tesla experts in robotics and AI I spoke with at the event were impressed with Optimus, the code name for the Tesla Bot. And Tesla engineers had the kind of fire in their bellies that bodes well for progress, especially when backed by Chief Executive Elon Musk's vision and Tesla's resources.
"I haven't worked this hard since grad school. But I love it," said one senior chip designer working on Tesla's Dojo technology for training the artificial intelligence systems behind Optimus' ability to navigate rooms and Tesla's FSD technology to steer cars. He wasn't authorized to speak to the media.
Seeing the Tesla Bot firsthand was impressive
I was impressed, too, compared with last year, when Tesla debuted Optimus as a few presentation slides accompanied by a robot-costumed human dancer.
I didn't applaud and cheer like many of the 1,000-plus people crammed onto a giant shop floor at Tesla's offices in Palo Alto, California, when the first Tesla Bot walked across the stage. I did appreciate the engineering achievement when seeing the computer-controlled biped pivot at the waist without toppling, powered by a computer on its chest, moved by a collection of gleaming cylindrical actuators, studded with status LEDs and cabled with a copper nervous system. I heard its cooling fans blow and, later, strained to heft the 42-pound block of copper and electronics that constitutes each 25-processor Dojo AI training tile.
It's not yet clear whether humanoid robots will someday do our shopping or take over from humans toiling on production lines, as Musk envisions. The controversial and outspoken billionaire estimates it'll be three to five years before the first Optimus goes on sale, an eternity in the technology world. But if his vision holds true, it could transform the world as profoundly as the automobile or smartphone.
"We wouldn't be surprised to see it becoming the primary driver of the stock price towards the end of the decade," New Street Research analyst Pierre Ferragu said in a report Monday. "Imagine Optimus as a startup today: It could be valued at several billion dollars, maybe even a few tens."
Research firm IDTechEx expects spending on service robots to nearly quintuple by 2032 to $70 billion.
To be sure, there's a long trip from prototype to product, as AI Day attendees were reminded when seeing a Tesla Semi and Cybertruck, two vehicles that have missed initial ship dates.
While Tesla's ability to deliver on its robot vision is even more up in the air, I'd recommend against writing the project off merely as Musk's utopian fantasy. Here's why.
Give Musk credit for creating nerd heaven
Musk has a talent for picking businesses that are difficult but attainable, for finding engineers up for the challenge, and providing them an environment where they expect to make a difference. They've managed to stay motivated even with Musk distractions like launching new companies and acquiring Twitter.
One who works at SpaceX, which has upended the rocket business the same way Tesla is shaking up the automotive industry, said rivals have tried recruiting her several times. They'd be cushy jobs, she said, but in practice, she wouldn't really get to do anything there. She wasn't authorized to speak to the media.
We might spend lots of money on laptops and smartphones, but Musk is looking for newer horizons. "You definitely want to see what's happening with Optimus, whereas a bunch of other technologies are sort of plateaued," Musk said at the AI Day event.
That message resonated with one engineer I spoke with at AI Day who's working on actuators, a key mechanism that moves Optimus' torso, legs, arms and fingers. Previously, he worked at Boeing, but he sees robotics as a new wave of innovation.
AI Day 2022 is explicitly set up as a recruiting event for engineers. One I spoke with, who works for a direct Tesla rival, showed up because of a recruiting email. Another who works on surgical robots was clearly impressed by the breadth and funding of Tesla's work in the area. Nobody was laughing off Tesla's Optimus.
Tesla Bot represents deep research
On stage at AI Day, a pre-Optimus prototype called Bumble-C shuffled, waved, pumped its arms and flexed at the waist. It was unspectacular by some standards, but also the result of less than a year's work. A series of engineers shared the stage with Musk to detail the research that's so far gone into Optimus. Among them:
- Tesla is designing its own actuators, tightly packaged sets of gears, motors, sensors and controls that act like human muscles. The company is drawing from the team that already designs Tesla cars' drive units. At AI Day, engineers showed Tesla's calculations on how it settled on a collection of six actuators for the whole robot that best optimized cost, ease of manufacturing, speed, torque, mass and efficiency.
- Tesla is drawing lessons from human anatomy, for example, employing a complex four-bar hinging mechanism similar to the human knee that adapts an actuator's differing needs for strength or speed depending on how far the knee is bent.
- Tesla is using the same AI technology behind FSD to guide Optimus, too. That includes the occupancy network, an AI system that converts camera input data into a 3D map of everything around a robot.
- Tesla is building its own custom data center technology, Dojo, to train the multiple AI systems that go into a vehicle or robot. So far, Tesla has built three cabinets of Dojo hardware and already it can handle video training data in ways impossible with the earlier approach using 72 of Nvidia's high-end A100 processors.
- For locomotion, Optimus combines a physics model that simulates a virtual robot with real-world sensor data that gauges how the robot is actually performing. The walking technology has grown more sophisticated: first steps were in February; Optimus' pelvis started rocking in July; swinging arm motion accompanying its strides starting in August; and Optimus' toes first lifted off the ground in September.
"The Tesla team is so far ahead and so confident in its pace of innovation that it doesn't worry about sharing details of what they do," Ferragu said.
AI is mandatory on Optimus
Optimus faces a challenge that Tesla cars do not: It requires AI for everything it does. Tesla can make a successful business selling industry-leading electric vehicles piloted entirely by humans then gradually develop autonomous vehicle technology.
Tesla in September expanded beta testing of FSD, its software designed to eventually lead to self-driving cars. 160,000 people in the United States and Canada have access, CEO Elon Musk said.
With humanoid robots, AI has to work from the start. And the AI challenge is harder with robots.
A quick primer on today's AI technology: Instead of programming computers with narrow, rigid if-this-then-that instructions, AI works by training a system to recognize patterns in vast quantities of real-world data. That lets AI systems handle much more complexity and make more nuanced decisions.
The tricky thing is that self-driving cars encounter a huge variety of situations. Even the same road intersection can be very different if it's raining, under construction or blocked by a stalled car.
Tesla is investing massively to overcome these challenges, slurping up 100 terabytes of video data reported by its cars every day, according to one engineer at Tesla AI Day. It also simulates various conditions to extend AI training situations beyond that real-world data.
But the variety of situations robots could encounter is vastly broader. Think of how different one home is from the one next door. Then compare those homes to businesses, sidewalks and farms.
Tesla's robot demonstration videos were set in a relatively narrow domain, though: its offices and research labs. And Musk said the company plans to first test them in its own "gigafactories." That could provide enough training data to give the robots a useful foothold.
Why build the Optimus robot?
Musk offered a somewhat murky motivation for why to build Optimus. SpaceX's job is to get humanity to Mars, and Tesla electric vehicles and batteries are to wean us from fossil fuels. Musk's explanation for Optimus seemed more opportunistic: Tesla has the expertise, so it might as well.
Musk hopes Optimus will "help millions of people," freeing humans from boring, dangerous and repetitive work. In his most starry-eyed moment, he said robots could lead to "a future of abundance, a future where there is no poverty, where people can have whatever you want in terms of products and services. It really is a fundamental transformation of civilization as we know it."
I wouldn't expect an end to poverty anytime soon. But today's Optimus is taking big steps, literally and figuratively, from last year's debut as just an idea.