Do androids dream of electronic sheep? Perhaps the answer lies in the world of modern robots.
Hans Peter Brondmo, former vice president of Google X and former CEO of Everyday Robots, says that at Everyday Robots, robots are dreaming all night, learning in a simulated world and "waking up" to pick up new skills.
Everyday Robots has a futuristic lab where roboticist Peter Pastor has configured seven industrial Kuka robot arms, figuratively called "Arm-Farm".
These robotic arms work around the clock, moving like a claw machine to a random location above the bin, and reaching for items such as sponges, Lego bricks, yellow ducks or plastic bananas in the bin. A camera above the bin captures every movement of the item and the robotic arm, helping the system determine whether the gripping was successful.
Through repeated training, the bot's success rate has grown from an initial 7% success rate to over 70%.
This result is exciting, but it is still not enough. It takes months to learn how fast to pick up a yellow duck, and a few hundred years won't allow a robot to learn to perform tasks in the real world.
That's why Everyday Robots has developed a cloud-based simulator that has created more than 240 million robot instances with built-in real-world physics models that simulate the weight and friction of objects.
Thousands of virtual robots perform tasks simultaneously with simulated cameras and simulated robotic arms, and experience millions of failures before applying their high-performing algorithms to real physical robots and training them in the real world.
Everyday Robots' achievements in the field of robotics learning are remarkable. The debate between "end-to-end learning" and "hybrid approaches" never stops, and Everyday Robots has made significant progress on both fronts.
In 2023, OpenAI launched ChatGPT, which also made Everyday Robots realize the importance of data, setting the next goal: to have millions of robots perform tasks in the real world, collecting enough data to train the final model to enable robots to go beyond the complex tasks that narrow the scope of training defined tasks.
It took seven years for Everyday Robots to go from picking up things to being able to walk around the Google building to do chores like wiping tables and sorting garbage, and was convinced that it was leading a future era of robots.
Hans Peter Brondmo talks to his mother on the phone every week, and his mother asks the same question every time, "When are the robots coming?" And he always replied, "Mom, it's going to be a while." ”
Sadly, Hans Peter Brondmo's mother passed away in 2021 and didn't wait for the arrival of robots, which also seemed to foreshadow the end of Everyday Robots.
Just two months after ChatGPT was launched, Google abruptly announced the closure of Everyday Robots, and only a handful of bot projects and members were transferred to DeepMind to continue research.
Hans Peter Brondmo 也从 Everyday Robots 离职，转向摄影领域，似乎有些心灰意冷。
Ten years ago, Google launched a robot acquisition frenzy, and companies such as Boston Dynamics were bought one by one.
Now, ten years later, Boston Dynamics has changed hands several times, and the glory of Google Robotics has fallen into a trough with the closure of Everyday Robots and the loss of many researchers.
The Garden of Eden for Scientific Research
In Google's prospectus to prepare for the IPO, there is an open letter that the two founders insist on attaching, which begins with this:
Google is not a traditional company. We don't plan to become such a company either.
The two founders were concerned that going public would undermine Google's independence and objectivity, and that it might forgo long-term interests because of short-term market demand. To guarantee the independence of management's decision-making, Google has created a "dual voting" structure: management's Class B shares have 10 times more voting power than publicly owned Class A shares.
At the end of the open letter, they emphasize:
Don't be evil. We strongly believe that as a company that does good for the world, we will be better rewarded in the long run – even if we give up some short-term gains.
The letter came to be known as the "Declaration of No Evil."
Google X was founded with this value in mind, developing and rolling out technologies that improve the lives of millions, if not billions. All of the lab's projects are collectively known as the "Moonshot," symbolizing the same effort as the moon landing program, aiming for a seemingly impossible goal.
Unlike mainstream workplace rules, Google X doesn't require employees to plan out in detail before making decisions or proposals, and CEO Astro Teller welcomes innovative and even crazy ideas from employees.
Astro Teller came up with the idea of installing a loudspeaker on a plant cart (a technology that rolls through the field to collect plant data) to broadcast motivational speeches to the plants. He doesn't care about other people's strange eyes, in his opinion, the value of ideas is not how good the results are, but whether they are innovative and can think outside the box.
"If teams take the time to explore and test stupid ideas like this, it can help 'break free' their minds and achieve results that might not otherwise be achieved," he said. ”
Of course, even Google X has to take into account the reality factor. There is an almost weekly review meeting where the members come together to discuss and reject the crazy ideas that the members have come up with. About 90 percent of projects will stop at proposing because they are too difficult or too expensive.
Once passed, the support given by Google X is also unimaginable.
Astro Teller had asked Google for $30 million for his project, and the other party gave him $150 million outright. Eric Schmidt, then CEO of Google, said, "If I give you $30 million, you'll come back next month and ask for $30 million." ”
At Google X, projects fail a lot more often than other labs. But Google X doesn't treat failed projects as a shame like traditional companies, shutting down the entire project and laying off the team when it fails.
Astro Teller has set up a project failure bonus to encourage people to take risks and comfort employees who have tried their best but have not succeeded.
Google X also held sharing sessions with employees at their suggestion, where they could share their experiences of failed projects and life changes.
Pursue innovation, encourage risk-taking, and embrace failure. In this atmosphere, Google X has created a pure research Eden.
Google has been exploring the combination of AI and robotics, with the goal of creating bots that can perform complex tasks. Everyday Robots, a universal learning robot project from Google X, is the product of this vision.
The research at Everyday Robots is progressing very well, and the team has developed an AI model that can be trained in a simulated environment, transferring the high-performing results to actual robot hardware for further testing and optimization.
While the complexity and unpredictability of the real world have led to a significant gap between how robots perform in simulation and reality, this hasn't stopped Everyday Robots' research.
These robots are becoming more capable of doing simple tasks like tidying up desks in a real-world environment, and this means that robots can effectively recognize humans and objects around them with the help of AI. Soon, these robots were given more responsibilities, busy in the Google building, handling all kinds of daily chores, and becoming real "special employees".
More than just chores, Everyday Robots extends the use of robots to the arts.
Catie Cuan, a former professional dancer, worked with engineers to develop an algorithm that could train robots based on choreographer preferences. In her free time, she dances with these robots and explores a new fusion of dance and technology.
Tom Engbersen from the Netherlands was intrigued, and together with Catie Cuan, he designed different sound effects for each joint of the robot. When the music mode is turned on, each movement of the robot creates a unique orchestral composition.
Hans Peter Brondmo and his mother have always believed that these robots will soon find their way into the everyday lives of ordinary people. But his mother didn't wait for this day, and the cruelty of reality doesn't stop there.
The clash of ideals and reality
Business black holes?
According to estimates by consulting firm Evercore, Google spent $650 million in R&D on Google X, but incurred a loss of $900 million.
Entrepreneurs and academics in Silicon Valley often criticize Google X, which "burns money", how can it eat up so many resources but delay in handing over a truly "phenomenal" product?
At the end of the day, Google X is more of an idealist innovation playground than a qualified business lab.
For example, the well-known Google Glass achieved a technological milestone, but ultimately failed due to high costs and a lukewarm response to the market's inability to commercialize.
Google Glass was officially released in April 2013, and the initial version of the Explorer Edition cost a whopping $1,500, or nearly 10,000 RMB. Needless to say, Google Glass has a lot of technology, but unfortunately the mainstream market doesn't buy it.
On the one hand, Google Glass's convenient shooting feature involves privacy issues, which has caused a lot of privacy controversy.
On the other hand, social acceptance of this novelty high-tech glasses is surprisingly low. According to a survey by mobile app company Bite Interactive, 38% of respondents said they would not wear Google Glass, even if the price was within their budget. Another 45% are concerned that wearing Google Glass will cause social embarrassment or unpleasantness.
Kevin Werbach, a professor of legal studies and business ethics at the Wharton School, also points out that "it may take years for Google Glass to come down to a level that is acceptable to the average person, and it may take longer for society to accept it." ”
Google Glass is just the tip of the iceberg of Google X's failed project. There are also projects such as Loon, a modular mobile phone Project Ara and an energy kite, all of which are unable to reach the market due to factors such as too complex or expensive technology.
Innovation can sometimes be like a gamble, and Google makes big bets, but not always right.
In 2015, Google established Alphabet as Google's parent company. At the same time, the original independent operation of Google X, DeepMind, Waymo and other departments were separated and put under the name of Alphabet as a subsidiary.
The purpose of Alphabet is to help Google focus on its lucrative core business without being held back by innovative projects that "burn money".
On January 20, 2023, Alphabet failed to weather the economic winter, announcing that it would lay off about 12,000 people, or 6% of its workforce. Everyday Robots was one of the few projects to be disbanded due to budget cuts, falling victim to this game of innovation and profit.
A common problem for large companies
According to Alphabet's 2023 financial report, the advertising business accounts for about 80% of Google's total revenue, cloud computing services contribute 10% of the total revenue, and the remaining 10% comes from hardware and other businesses.
Google's supremacy in the advertising industry guarantees a steady stream of huge revenues and reflects its heavy reliance on the advertising business. Even cloud computing services, a new profitable growth point, remained in the red until 2023, when they finally made their first profit.
However, there are really many failure cases in the market where there is a single profitable project and is highly dependent on the income of a specific business.
The most famous case is Nokia, once the world's largest mobile phone manufacturer, which had no rivals in the feature phone market, but due to the failure to transform smartphones in time, the market share was rapidly lost, and it was finally acquired by Microsoft in 2013 and withdrew from the mainstream mobile phone market.
Another classic case is Yahoo, which once also relied on portals and search engines to become an internet giant, and its advertising business was also its main source of revenue, but it was acquired by Verizon and restructured into Oath in 2017 after losing ground under Google's offensive.
In the past, how Yahoo fell is still vivid, and now it is possible to follow the old path of Yahoo, and Google is also very anxious, otherwise it will not stop the research when its own robot momentum is good, and turn around and follow OpenAI to start making large language models.
There is a peculiar phenomenon at Google. On the one hand, all kinds of innovative projects emerge in an endless stream, and only the industry can't think of it, and Google can't do it; On the one hand, Google does not have a trump card project that can stand on the top of the mountain in areas other than advertising, and it is always relegated to the laggard.
The problem of profitable open source is tightly stuck in Google's neck, but the disaster is not alone, and Google is starting to lose people.
Sheryl Sandberg, Google's VP of Advertising & Commercialization, left for Meta, Thomas Kurian, VP of Cloud, left for Oracle, and YouTube CEO Susan Wojcicki ended a 25-year career at Google.
Many of the employees who participated in Google's robot moon landing project went to other companies: Yan Mengyuan went to OpenAI, Eric Jang chose the startup 1X, and Yao Lu, one of the initiators of Deepmind RT-1, RT-2, RT-X and other projects, is now the chief researcher of Nvidia......
A senior operations engineer at Google wrote a resignation report on his blog, looking back on his nine-year journey at Google:
At first, she was attracted to technology by Google, but after her continuous promotion, she felt more and more stressed due to the company's bloated organization, tedious and meaningless meetings and repetitive work, marginalization, and frustrated innovation, and finally chose to leave Google.
This also reveals a larger problem: the expansion of the company inevitably brings complex hierarchies and inefficient decision-making, and Google, which wants to rely on dual voting rights to get rid of this problem, ends up suffering from the common problem of bloated and rigid large companies.
A former employee of Everyday Robots said in an interview that it was difficult for the company to determine the company's mission. Teams can't decide whether they're aiming for an advanced research project or a commercially viable product.
Google, which once had a clear goal, is now lost at the intersection of innovation with a vague mission.
Google's innovation is lost
What's the problem?
Back in 2010, Google founders Larry Page and Sergey Brin decided to set up a new division to study the "moon landing":
Technology that sounds distant and sci-fi may one day make the world a better place.
Google X is this bold experiment, and it has indeed incubated a lot of "black technology" that sounds like science fiction.
For example, the famous Google Glass, after wearing it, can directly release a virtual screen in front of your eyes, which can be manipulated by gestures, voice, etc., and is also equipped with functions such as bone conduction.
Another example is the 280 Earth project, which aims at carbon emissions, where the amount of carbon dioxide in the atmosphere is too high, and the carbon dioxide can be used to produce fresh water, and the waste heat in the middle can be used to power the next stage.
Google set up Google X to cram all the highest-risk, lowest-short-term return innovation projects into one place and operate in the form of an "external brain".
As early as the beginning of the IPO, Google's two founders were worried that the short-term interests demanded by shareholders in the future would conflict violently with the long-term investment required for scientific research, causing Google to lose its original intention and abandon long-term projects, and stop innovating and only making quick money.
In order to avoid this situation, there is a special Google X, which is both the "external brain" and the original intention.
Google used to have a "20% innovation time" policy, which meant that employees could dedicate one day a week to work on projects other than their own work. This innovative regulation has led to great products such as Gmail, AdSense, Google News, and GTalk.
But this policy has since been repealed and was considered "dead in name" by internal employees back in 2013.
At that time, employees had to apply for this benefit if they wanted to take advantage of it, and it was becoming more and more difficult to get approved, and there was an internal employee efficiency testing system to determine whether employees were 100% committed to their jobs.
In fact, the abolition of "20% innovation time" is also the inevitable result of Google's pursuit of efficiency and standardization after the expansion of its scale, but Google has not given up the pursuit of innovation. In a sense, Google X is a stand-in for this policy, taking on the mission of innovation.
But judging from the performance in recent years, Google has not only failed to maintain the previous innovation atmosphere, but also moved towards a new innovation trap.
Eric Schmidt, the former CEO of Google, let himself go in a speech at Stanford not long ago, saying that "Google will lose to startups like OpenAI in the AI race because of Google's remote work policy, which causes employees to not work hard enough." ”
This is a microcosm of Google's innovation trap. Executives are reluctant to relax working hours, and want to force employees to focus on their work by tightening management policies, hoping that this will improve work efficiency and allow employees to produce a steady stream of innovative products.
Google X's design kitchen
However, history tells us that the most innovative and energetic time for an enterprise is the start-up period, and only at this time is the enterprise truly innovation-oriented; Once a business scales, it has to face the trade-off between innovation and a profitable core business.
History also tells us that history is always strikingly similar, and that human beings always repeat the same mistakes.
Nokia, Yahoo, and BlackBerry have all missed the opportunity to transform because they were too conservative and not innovative enough, and were abandoned by the cruel market. But Google still has the luck that "change isn't coming so soon", and coupled with short-term financial pressures, it is gradually falling behind in innovation.
The embodied intelligence craze has come, does Google still have a chance?
According to CB Insights, total investment in the global robotics industry reached $8.9 billion in 2022, an increase of 44% from 2021. Among them, embodied intelligence startups such as Agility Robotics and Boston Dynamics have received large-scale financing, and the global robotics market is expected to exceed $40 billion by 2025.
In the past two years, the development of embodied intelligence can be said to have advanced by leaps and bounds, and major companies and scientific research teams have made great breakthroughs.
Nvidia is one of the leaders. For example, the GR00T project launched at the GTC developer conference this year was previously rated as one of the "Top 10 NVIDIA Projects in 2023". In addition to the full-stack robotics platform, NVIDIA also announced an acceleration program for humanoid robot developers, which aims to further accelerate the rapid development of embodied intelligence.
Not only NVIDIA, but also the 1X company invested by OpenAI is not far behind, launching a brand new humanoid robot, NEO, and releasing their own model of the world.
Feifei Li's team at Stanford University has also made a major update to the housekeeping robot and proposed a new ReKep technology. At the same time, World Labs, founded by Li Feifei, announced plans to build a "big world model" that will allow AI to perceive, generate, and interact in 3D worlds.
Chelsea Finn's team, another pillar of embodied intelligence at Stanford University, partnered with DeepMind on the ALOHA project to launch the Mobile ALOHA stir-fry robot earlier this year. RECENTLY, THEY INTRODUCED AN UPGRADED VERSION OF THE ALOHA 2 ROBOT, WHICH IS CAPABLE OF PERFORMING MORE FINE MOVEMENTS. In addition, the Chelsea Finn team also proposed a new algorithm for bidirectional decoding of BID.
However, not all companies have survived the wave of embodied intelligence.
An employee of Everyday Robots once said in an interview, "The company made premature promises about useful and affordable robots that could not be kept, and I hope Everyday Robots will not make the same mistake." ”
Everyday Robots, which once adhered to the principles of the "moon landing" program, fell ahead of the craze for embodied intelligence, but at Google, its mission doesn't end there.
The robotic arm that used to "catch dolls" at Arm-Farm is now in action in DeepMind's new project. The robotic arm played back and forth with a human player, becoming the first learning robot agent to reach the level of a human amateur player in a table tennis match.
Google still has a chance to win in this wave of embodied intelligence, but only if it makes something amazing enough to set it apart from its competitors.
When you win, whatever you say, it's right.
There is a famous saying in the customs study, "After entering the customs, there will be a great Confucianism to defend me." "As long as your strength is strong enough, someone will naturally defend you and stand in line for you. In fact, Google did not lose in management policy, but only in the courage to innovate.
Innovation needs to grow wildly rather than a defined framework, and dancing in shackles will inevitably fail to achieve qualitative breakthroughs.
Young people who started out tinkering with search engines in garages were able to storm the Internet giants with just one heart. If Google doesn't want to miss out on this wave of embodied intelligence, it should rethink its timid status quo and regain the courage to subvert itself and take on the challenge of rising stars. Leifeng Net, Leifeng Net