|
|
|
|
|
|
|
|
Super-Capable Robots to Share Human Labor
Technology Briefing |
|
TranscriptSimplistically, we recognize that the Industrial Revolution totally changed the trajectory of civilization. Between its beginning, around 1770, and the year 2000, real U.S. GDP per capita soared by roughly 40 times. This involved more than one quantum leap. In fact, it happened in five successive waves we refer to as Techno-Economic Revolutions. In the 2020s, we’re in the midst of the second phase of the Fifth Techno-Economic Revolution, which is based on “information processing.” As explained in prior issues, this current stage in the revolution is being enabled by progress in hardware, software, and networks, which are finally making artificial intelligence cost-effective. That means technology can do more jobs formerly performed by humans. In turn, that indicates that rapid growth could result from AI automating tasks in the production of goods, services and ideas, with the latter being especially important for explosive growth. Those who have been following the progress of Generative and Analytic AI, already understand how these technologies are likely to super-charge “knowledge work” and “idea generation.” But let’s not forget about automating the production of goods and performing services. For that to happen, it will take great hardware as well as great software. It will involve super-capable robots that can easily substitute for a major share of human labor. Large language models by themselves aren’t enough. In economies with rapidly growing literate populations, such as the United States (circa 1970) or China (circa 1990), a breakthrough like this would threaten to unleash mass unemployment. However, in the “Peak Human” world beyond the 2030s, such automation is probably the only way to accelerate or even continue economic growth. But here’s the problem: robots need abundant data from physical interactions to learn how to operate in the physical world, data which are costly and time-consuming to produce in labs. To-date progress has been limited, with robots only able to perform simple behaviors in restricted settings. This is all neatly explained in an IEEE Spectrum essay titled, “The Global Project to Make a General Robotic Brain,” by Google scientists Sergey Levine and Karol Hausman. “If the abilities of each robot are limited by the time and effort it takes to manually teach it to perform a new task, what if we were to pool together the experiences of many robots, so a new robot could learn from all of them at once? We decided to give it a try. In 2023, our labs at Google and UC Berkeley came together with 32 other robotics laboratories in North America, Europe, and Asia to undertake the RT-X project, with the goal of assembling data, resources, and code to make general-purpose robots a reality. The RT-X project aims to collaboratively pool data and resources at scale, with the goal of creating versatile, general-purpose robots that can operate effectively beyond limited lab settings. To our surprise, we found that the inclusion of the multirobot data improved the Google robot’s ability to generalize on such tasks by a factor of three. This result suggests that not only was the multirobot RT-X data useful for acquiring a variety of physical skills, but it could also help it better connect such skills to the semantic and symbolic knowledge in vision-language models. These connections give the robot a degree of common sense, which could one day enable robots to understand the meaning of complex and nuanced user commands like “Bring me my breakfast” while carrying out the actions to make it happen.” In a recent report titled, “Humanoid Robot: The AI Accelerant,” Goldman Sachs analysts upgraded their optimistic outlook for the emerging humanoid robot market. James Pethakoukis of the American Enterprise Institute highlighted three key points from that report:
In a 2022 report previously cited in Trends, Goldman Sachs described their “Blue-sky scenario” as one where the humanoid robot sector would be a $150 billion market by 2035 with an average price tag of under $20,000. Of course, that assumed that aggressive costs and functionality targets were met and other barriers “to wide public acceptance could be completely overcome.” This optimistic market penetration would fill from 48% to 126% of the factory labor gap, and as much as 53% of the elderly care-giver gap. The “Blue-sky case” implies a demand for 3-to-5 million humanoid robots in the USA alone by 2035. An updated forecast from Goldman Sachs implies a market that’s 50% larger than the 2022 estimate with 16% of sales coming from industry and 84% Tesla Bot development timeline vs Boston Dynamics Atlas coming from consumers. A lot would have to go right to make this scenario a reality, but even the updated base case would be a game-changer for the economy, while being less disruptive. Now, as in 2022, Tesla’s progress at developing a humanoid robot is seen as a reason for this optimism. On December 13, 2023, Tesla unveiled the Tesla Optimus Bot Gen 2 in video format which represents another successful milestone on its aggressive timeline. Notably, Tesla launched its Gen 1 product in March 2023 and demonstrated the end-to-end AI learning capabilities in June, showing a much faster product iteration and R&D pace than Boston Dynamics and its other peers. Tesla released a video in Sept. 2023 to explain the “end-to-end AI process” used in Optimus. Tesla says end-to-end AI is completely different from prior rule-based control, meaning the software system itself can execute the task all the way from original commands and scenarios to final outputs using self-generated AI rules instead of software engineers’ pre-programmed rules. In the video, Tesla’s Optimus completed higher-level end-to-end operation from scenario observation to task analysis. These two processes are completed by “robot LLMs” which have simulation and analysis abilities, as well as image, text, and video reasoning. The lower-level end-to-end operation includes simulation and final execution, where the robot’s AI needs to arrange the workflow of movements (using simulation) and then turn those into the robot’s physical movements that we see. What’s the bottom line? Service and manufacturing robots are moving to create a cost-effective bridge between artificial intelligence and the physical world. The question is still precisely when, where, and how. Fortunately, managers, investors, consumers, and policy makers all still have time to evaluate and exploit this enormous opportunity. Given this trend, we offer the following forecasts for your consideration. First, the shortage of U.S. manufacturing workers will reach at least 2 million by 2030, creating demand for a million or more humanoid factory robots. Manufacturing and warehousing are the areas of the economy where humanoid robots will appear first because these environments are self-contained, the work is standardized, and the cost-benefit equation is easy to analyze. Furthermore, tireless, reliable robots will be especially attractive for dangerous applications as well as use in green-field plants optimized for robots. Trends in plant safety standards and re-shoring encourage robot adoption. The biggest questions center around when cost decreases and functionality increases will converge to make the payback “good enough” to justify widespread corporate adoption. Fortunately, even at $150,000 per unit, humanoid robots will prove cost-effective for serious factory work once they can reliably perform “real jobs.” Ideally, many factory jobs will remain unsuitable for robots, permitting humans to work synergistically alongside them. Second, the market for household and service industry robots will develop later because of greater price sensitivity and higher levels of “job variability.” Even for humans, working in homes and service venues requires greater flexibility and adaptability than factory work. Not only do homes and service establishments typically lack the predictability of factories, but service employees are often required to do a wider range of tasks and deal with a greater level of ambiguity. To address this situation, solutions such as Tesla’s “end-to-end AI” and the multi-robot learning enabled by Google’s RT-X project, are aimed at rapidly overcoming this learning challenge. Even so, the consensus of experts is that “reliable and general household robots” are at least a decade away. Third, over the coming decade, no area of technology will advance more rapidly than training robots. Suddenly, a wide range of now approaches has emerged from industrial and academic labs around the world. The printable issue shows screenshots from videos researchers at Carnegie Mellon are using to train robots to perform household tasks. The idea is for robots to learn where and how humans interact with different objects by watching videos. Previous methods of training robots required either the manual demonstration of tasks by humans or extensive training in a simulated environment. Both are time consuming and prone to failure. Using this approach robots can learn a new task in as little as 25 minutes. Starting with 4000 hours of videos, the researchers are busy training robots to handle common household tasks that are generalizable to new environments. As industry-wide experience accumulates, performance should rise exponentially. Fourth, by the late 2030s, complex mechanical actuators will be replaced by simple, low maintenance artificial muscles. Researchers around the world are working on energy-efficient, long-lived muscles for robots. Until recently artificial muscles involved slow response times, high voltages, or both. However, recent breakthroughs are solving these and other problems. So, by the time robots are commonplace in homes they are likely to look and sound a lot more like humans than any of today’s prototypes. And, Fifth, despite the myriad competitors working to address the humanoid robot market, Tesla will perform extraordinarily well. As we’ve seen with SpaceX, X, and even Tesla’s car business, Elon Musk has a unique talent for redefining or creating businesses. At Trends, much of our skepticism about Tesla grew out of doubts about the viability of the battery-based EV industry. While recent results from Ford, GM, Mercedes, and others have validated our assessment of the EV opportunity, Tesla’s innovations have enabled it to gain share in a failing market. And as the EV opportunity shrinks into the rearview mirror, we expect Tesla to leverage its work with the self-driving “Auto Pilot,” as well as batteries, sensors and (most importantly) AI, to transform itself into a leading “robot company.” And by having its own in-house automotive plants, it’s especially well-positioned to pioneer humanoid robots in manufacturing applications. |
|
Comments
|
|
|
|
|
|
|
|
|