Musk’s Optimus robot disappoints, but his ambition might just transform the field anyway
Well, now we’ve seen it: the Optimus humanoid robot that Elon Musk has been talking about for the past year. On Friday, at Tesla’s “AI Day” —actually a misnomer, since the event was held in the evening California time—the billionaire introduced a working prototype of the robot, which he envisions will one day do a wide variety of helpful tasks in people’s homes.
Although Musk has spent a lot of time hyping the Optimus—saying it will help to usher in an age in which robots and A.I.-enabled software perform most of the economically useful work now done by humans—just before he unveiled the new robot he tried to tamp down expectations for what people were about to witness. After all, he told the audience, last year all Tesla had managed to show off with regards to the Optimus was a human dancer dressed like a robot. Now they at least had a working machine, he said. “Compared to that it is going to be very impressive,” he said. And then one of Musk’s assistants noted that it was the first time Tesla was trying the bipedal robot without any kind of mechanical support to ensure it did not topple over.
The robot—this version, which Tesla calls Bumble C, was built largely using motors, acutators and other parts from third-party vendors and using A.I. computer vision and navigation software adapted from Tesla’s own Autopilot advanced driver assistance software—stepped hesitantly onto the stage and waved to the audience. It even did a little dab-like dance. “The robot can actually do a lot more than what we just showed you,” Musk said. “We just didn’t want it to fall on its face.” He then showed videos of the robot performing a few tasks: carrying boxes in a warehouse and office, placing the box on a desk, picking up a watering can and watering some plants, and picking up some metal parts in a Tesla factory.
Musk then showed off a slightly slicker-looking version of the Optimus that had been built using Tesla-designed parts. “It wasn’t quite ready to walk, but it should be ready to walk in a few weeks,” he said. It dutifully was wheeled out on a kind of stand and waved. Musk said the Optimus would ultimately be made in “millions of units” and would cost “much less than a car”—less than $20,000, he said.
So how impressed should we be with the Optimus? Most roboticists were underwhelmed. On Twitter, Cynthia Yeung, a roboticist at Plus One Robotics, which builds software for logistics robots, wrote: “None of this is cutting edge. Hire some PhDs and go to some robotics conferences @Tesla.”
In the technical journal IEEE Spectrum, the publication’s senior editor and long-time robotics-watcher Evan Ackerman wrote:
“While there’s absolutely nothing wrong with the humanoid robot that Musk very briefly demonstrated on stage, there’s nothing uniquely right, either. We were hoping for (if not necessarily expecting) more from Tesla. And while the robot isn’t exactly a disappointment, there’s very little to suggest that it disrupts robotics the way that SpaceX did for rockets or Tesla did for electric cars.”
Will Jackson, the founder and CEO of Engineered Arts, a company known for its humanoid robots, said he was impressed with Tesla’s A.I. and the company’s engineering abilities, but he too was underwhelmed with what Musk demonstrated. “The first two prototypes shown at the unveil are decidedly lacking,” he said in an email. “They are very similar in concept to Honda’s Asimo robots, the development of which is now abandoned.” He said that mechanically, “there’s no novelty here.” But Jackson tempered his remarks by saying that Tesla had a capable engineering team and would likely improve the robots significantly in future updates.
Beyond concern about the hardware, the demos also didn’t make it clear how good Tesla’s A.I. software for the Optimus actually is. It was particularly unclear how capable the robot really would be in a household, which is where Musk has talked about the Optimus ultimately being deployed. Did the robot need separate training for watering plants than it did for moving the heavy metal parts in Tesla’s factory, for instance? What if it had to place the box in a different office that was not configured the same way? How well could it avoid bumping into people who walked into its path? How dexterous are its hands? Can it fold clothing? Can it put away the dishes? Can it mind your children or pets? None of those questions were answered in the demos.
As for whether there’s a market for the Optimus in factories and warehouses, which is where Musk has said the Optimus will first see use, there are plenty of other non-humanoid robots that already function perfectly well in those settings. It’s not clear what a slightly wobbly bipedal robot would add in those markets.
A few experts told me they were impressed with Tesla’s ambition: a humanoid robot at less than $20,000 could be game-changing, according to Pieter Abbeel, the UC Berkeley professor who co-founded Covariant, among the world’s top companies for A.I. software to enable robots to perform a diverse array of tasks. He said that a $20k robot would give buyers a machine that could be used in a vast array of settings—almost anywhere that a human can go—at a price point that is less than the cost of a single industrial robotic arm. He said the key would be ensuring the robot was as reliable as current industrial robots, with lifespans of at least 10 years and relatively minimal maintenance requirements.
“Let’s just say I very much hope that to become true soon,” Abbeel said in an emailed response to my questions. “I would love to start doing our AI research at Berkeley (and possibly AI robotic deployments at Covariant, but speed matters a lot there, and speed wasn’t super-clear) with such a robot.” Abbeel said Tesla’s approach—which sees improving the A.I. software for the Optimus, more than any hardware innovation, as the key to bringing it to market, was the right one.
Meanwhile, Jackson, from Engineered Arts, said he was surprised that Musk seems so focused on replacing people in factories and industrial settings—as opposed to building a machine for human interaction. The “true killer app” for a humanoid robot are people’s desire to interact with it, he said.
The real impact of the Optimus may not be what the robot itself does, but its impact on the mass adoption of humanoid robots in general. This is the “Elon Musk effect.” Sure, Musk hypes the technologies he works on. But—as he has with electric vehicles, space-based businesses, the hyperloop, and even brain-computer interfaces—his interest tends to act as a magnet for others, drawing serious attention and serious money to entire product categories and business models. So while the Optimus may not be walking your dog any day soon, expect to see a lot of other humanoid robots begin to make their way to market in the coming years.
It’s not too late to join me this Thursday—October 6—at 12 p.m. ET for what promises to be a fantastic virtual round table discussion on A.I. “Values and Value.”
As companies tackle the ethical problems that arise from the widespread collection, analysis, and use of massive troves of data, join us to discuss where the greatest dangers lie, and how leaders like you should think about them.
- Naba Banerjee, Head of Product, Airbnb
- Krishna Gade, Founder and CEO, Fiddler AI
- Ray Eitel Porter, Managing Director and Global Lead for Responsible A.I., Accenture
- Raj Seshadri, President, Data and Services, Mastercard
You can register to attend by following the link from Fortune’s virtual event page.
And, if you want to know more about how to use A.I. effectively to supercharge your business, please join us in San Francisco on December 5th and 6th for Fortune’s second annual Brainstorm A.I. conference. Learn how A.I. can help you to Augment, Accelerate, and Automate. Confirmed speakers include such A.I. luminaries as Stanford University’s Fei-Fei Li and Landing AI’ Andrew Ng, Google’s James Manyika, and Darktrace’s Nicole Eagan. Apply to attend today!
A.I. IN THE NEWS
The Biden Administration Proposes ‘A.I. Bill of Rights’ — In a document it called a “blueprint” for an A.I. Bill of Rights,” the Biden Administration’s Office of Science and Technology Policy laid out five broad principles that it says should govern all uses of A.I. It said users should be protected from automated systems that are unsafe or that use inappropriate or irrelevant data; that A.I. systems should not discriminate against protected classes of people; the inidividuals should have agency over how their data is used and have protection from “abusive data practices;” that people should know when a decision that impacts them is being made by an automated system and understand how and why it impacts them; that people should be able to opt out—“when appropriate” the document says—of automated decision-making and have access to a human who can be an alternate decision-maker or help remedy any errors. That all sounds good, but as Wired noted in its coverage, the document has no enforcement mechanism outside the federal government’s own A.I. systems. As with other statement of A.I. principles, Wired writes, “their tenets are usually directionally right, using words like transparency, explainability, and trustworthy, but they lack teeth and are too vague to make a difference in people’s everyday lives.”
U.S. plans more limits on Chinese A.I. and supercomputing technologies. The New York Times reports that the Biden Administration is expected to announce sweeping further restrictions on the sale and export to China of semiconductor technology used in high-performance computing. It also plans to restrict U.S. companies, especially those working in defense, from purchasing even mid-range and older semiconductors, including memory chips, made in China, the paper said. The new rules could be announced as soon as this week.
U.S. lawmaker calls Stable Diffusion an “unsafe model” and urges export controls. Rep. Anna Eshoo, a California Democrat, wrote to the White House’s Office of Science and Technology Policy and the U.S. National Security Agency urging them to consider blocking the export of the text-to-image generation A.I. software Stable Diffusion, which she called “unsafe.” She said that Stable Diffusion is a “dual-use technology” that should be subject to strict controls because the software makes it very easy to generate violent and pornographic imagery as well as minsinformation and disinformation. Unlike some rival software, such as DALL-E from OpenAI, Stable Diffusion, which is produced by startup Stability A.I., has no content filters on either the data used to train the A.I. model or its outputs.
Meta releases software to make it easier to switch between Nvidia and AMD chips for A.I. applications. Meta, the social media giant, has created a set of tools for A.I. developers that makes it easier to deploy A.I. applications written in the popular programming language PyTorch across graphics processing units (GPUs) made by different vendors, such as Nvidia and AMD. In the past, A.I software once it was trained and deployed to make predictions—a task known as inference—worked best if optimized for just one set of GPU hardware or another. The new tools aim to make it easier to achieve the same performance while being hardware agnostic, according to a Reuters story. The software is free and open-source.
Bruce Willis denies giving company rights to create deepfake of him. Deepcake, a Delaware-registered company whose CEO and founder is, according to her LinkedIN page, based in Tbilisi, Georgia, made headlines last week when it said it had been given permission to use actor Bruce Willis’s image to create a deepfakes of Willis for use in future films and advertising campaigns. Willis had previously announced his retirement from acting due to a cognitive disorder known as aphasia, that makes language and speech difficult. But Willis’ manager issued a strong denial of the story, saying that Willis had not sold rights to his face or image. My Fortune colleague Alice Hearing has more here.
EYE ON A.I. TALENT
Productivity software startup Notion has hired Rama Katkar to be its first chief financial officer, according to a story in tech publication Protocol. She was previously Instacart’s vice president of finance.
Splunk, the network data and data analytics company, has hired Gretchen O’Hara to be its channel chief, according to trade publication CRN.com. She was previously at Microsoft, most recently serving as vice president of U.S. AI and Sustainability Strategy.
EYE ON A.I. RESEARCH
Battery breakthroughs brought to you by A.I. A team of scientists at Carnegie Mellon University in Pittsburgh have used a machine learning system called “Dragonfly” to design new kinds of electrolytes for lithium-ion batteries that would allow them to charge far more rapidly than existing models can, British newspaper The Independent reported.
The A.I. system found six electrolyte solutions that outperformed existing ones, with the best yielding a 13% performance improvement for the battery cell compared to commercially-available ones, the paper said.
FORTUNE ON A.I.
Elon Musk plans to put an Optimus robot in every home. On the cusp of Tesla’s ‘AI Day’, the question is whether this is the year he finally unveils one—by Jeremy Kahn and Christiaan Hetzner
Is it time for schools to start teaching "prompting?" The output of many of today’s most powerful A.I. systems can be refined through what are known as prompts. These are instructions about the kind of output the user is looking for. In the case of a text-to-image generator, the prompt might be something like, “a futuristic restaurant on Mars, vivid colors, photorealism”; in the case of some coding A.I. systems, it might be an instruction like “Create the polynomial function two times x squared plus four times x plus seven” or “Calculate a measure of center for Bitcoin's price over the past 10 days.” (Both of those are real examples from GitHub’s CoPilot and Salesforce’s CodeGen respectively.) In other cases, as with GPT-3, the prompt might be a sample of writing that the user wants the A.I. system to continue in the same style.
This is being referred to as “prompt engineering” and there are people trading—and even selling—guides for how best to prompt to models such as DALL-E and GPT-3. But “engineering” may be the wrong word. At least for now, there’s far more art to prompting than science.
A lot of education today is focused on techniques for doing things—how to write, how to draw and paint, how to play music, how to code, how do statistics. But what if in the future—and it seems like that future is rapidly approaching—the actual steps to execute a task become somewhat irrelevant to the person who wants to complete that task. Instead, the key to completing any task successfully will be learning how best to prompt an A.I. system, in plain natural language, to produce the desired output. Creativity and skill will be all about prompting, not about execution. Will schools need to teach prompting as a basic subject?
Our mission to make business better is fueled by readers like you. To enjoy unlimited access to our journalism, subscribe today.