Can AI Learn to Fly?

Adam Jones: Robots have been around for a while. General Motors first introduced industrial robots in a body shop ina New Jersey factory back in 1961. Yes, once upon a time cars were made in New Jersey. In the first century AD, Hero of Alexandria described robotics in great detail in his book, “Automata,” around the same time he invented the steam engine.

Pre-AI, robots executed highly predictable and repeatable tasks with limited use cases, a lack of awareness and adaptability. Post-AI, robots are on a path to execute more diverse and open-ended tasks, enabling these electromechanical wonders to escape from the factory and proliferate throughout the physical world, becoming part of our daily lives.

In Volume 2 of the Morgan Stanley Robot Almanac, we go deeper into AI-enabled robots, showing you the difference between a Foundational Model and a World Model, and the exquisite interrelationships between the two. We show you the difference between Large Language Models and Vision Language Action Models to help understand the value of historic data, or “priors,” versus the value of arising data, or “context,” continuously improving the predictive decision-making of your robot.

The multiple ways that robots learn through teleoperation, simulation, and two- or three-dimensional videos. And how video game engines play a potentially critical role in narrowing the sim-to-real gap for robotics.

Nvidia enables robots through three critical compute domains spanning data center, simulation, and inference at the edge. Tesla, a robot company, is building a distributed inference cloud that interconnects its cars and humanoids into a giant mobile, swarming, intelligent data center. An extraordinary wide range of firms are trying to capture real-world data to train robot models across thousands of use cases.

In the near future, your face and upper body will be increasingly adorned with a range of sensors. We expect the number of Meta Glasses in use will double the number of Teslas on the road in just a few years, using millions of exaflops of edge compute by 2050.

Is it any wonder that robotics are attracting an enormous acceleration of venture funding over the past two years? Our Robot Almanac tracks hundreds of public and private companies driving the birth of this

new Industrial Revolution, including deeper dives into key foundational model players such as Google DeepMind, Meta, Apple, Physical Intelligence, and Skilled AI.

Competition for AI supremacy between China and the United States is a powerful force vector of innovation. And while the U.S. may have an apparent lead in AI compute and AI models, China is in a dead sprint on the physical side, rapidly converting conventional manufacturing processes to make AI-enabled robots at high volumes, installing more industrial robots than all other countries combined.

China will soon lead the world in robot density, if they haven’t already. China is prioritizing scientific education over its chief economic rivals, continuing a pattern where the United States invents a novel

technology only for China to scale it to broad adoption and economic dominance. In 2024, China installed 1.5 times the solar capacity of the entire existing installed base of the United States. Let that sink in.

But while the United States needs China for rare earth magnets, China needs the United States for its insatiable consumer and bleeding-edge AI technologies, keeping the rivalry in a rather delicate, if at times precarious, balance.

Rare earths. Some of the names of the metals may be hard to pronounce, but they’re even harder to live without. The use cases? Let’s start with every physical sector: autos, A&D, robots, boats, C.E., energy, healthcare. Basically, everything that moves.

And guess who controls around 60% of the world’s rare earth mining? And that’s just the mining. Who refines over 90% of the world’s rare earth metals? And you better get used to it because it takes about two decades to spool up a new rare earth mine. Our Global Robot Model, or GRoM, estimates the rare earth magnet content across all key categories of robots, projecting 1.7 megatonnes of magnets by 2050.

In 1995, astronomer and planetary scientist Carl Sagan made a disturbing prophecy in his book, “The Demon-Haunted World,” about a future United States where nearly all of its manufacturing industries slipped away to other countries, leading to adverse geopolitical and societal outcomes.

Well, we may not be there quite yet, but judging by the last 80 years of U.S. manufacturing slipping from 28% of GDP to less than 10% of GDP today suggests we’re well on our way. The intersection of AI and the physical economy offers the chance to disprove Sagan’s prophecy, and we’re all part of the story, however it ends up.

The robots are coming. Stay human.

Adam Jonas: In 1921, Czech playwright Karel Čapek published R.U.R., a futuristic story where human-shaped machines did all the world’s work. Life was wonderfully abundant until the machines started to make certain demands. It’s the first time the word “robot” was ever introduced to popular culture.

People have been trying to make human-shaped robots for centuries, with more intelligent attempts in the late 20th century. What’s changed today is that we are witnessing the intersection of generative AI and robotics working together, where data and compute accelerate a robot’s ability to learn, enabling more advanced manipulation and perception of the real world and feeding the AI flywheel in a recursive loop.

Humanoid development is driven by the collision of two powerful and orthogonal forces: labor shortage and AI, accelerating societal adoption and economic paybacks.

What’s the addressable market for humanoids? Well, we’ve got four billion workers in the global labor market at an annual average wage of $10,000, driving a $40 trillion global TAM. Humanoids may be particularly useful for boring, dangerous, and repetitive tasks that require a range of flexibility.

We estimate the global installed population of humanoids will exceed one billion units by 2050, driven largely by Asia.

Our Almanac outlines the trade-offs between specialized robots and humanoids, and the differences between LLMs that train on internet data and humanoid foundation models that train on real-world data from physical environments.

Manufacturing of the robot probe cannot be separated from the data collection of the probe. You need to make the probe to collect the data, to improve the probe, to collect more data, to improve the probe. I think you get the idea here.

Data defines the software. Software defines the hardware. Hardware defines the manufacturing.

A recurring theme in our Robot Almanac is the dominance of China over the rest of the world. Humanoids are no exception. We’re following over a dozen Chinese firms that have unveiled robots to date. There may be over a thousand firms below the surface in China.

Our team’s tracking of patents suggests an increase in humanoid development and a total domination of patent filings in China versus all other countries.

Beijing has made humanoids a national security priority for the PRC in the country’s 15th Five-Year Plan, unveiled this year. These efforts address a host of national security risks, including China’s unique demographic challenges.

China is nurturing a culture that celebrates robotic competition, showcasing the ever-improving speed, strength, and agility of the humanoid form factor. China’s mature and vertically integrated auto industry is aggressively testing, developing, or manufacturing its own humanoid robots.

The Trump administration has acknowledged the strategic importance of establishing a redundant domestic supply of humanoid robots, while major U.S. technology firms are entering the arena directly or as enabling partners.

Our Robot Almanac draws intriguing comparisons between human biology and humanoid morphology, including joints, metabolism, the nervous system, sensory capability, and skeletal structure.

Our proprietary robot model explodes the humanoid into its component parts, including motors, reducers, screws, bearings, vision systems, chips, and other sensors. Our Robot Team has applied values to these components that help inform our identification of key stock exposures.

Working closely with our China Robots Team, the Almanac pays close attention to humanoid component suppliers across actuators, as well as tracking key sensor content across a range of humanoid companies.

Morgan Stanley’s Global Semis Team is highly focused on the humanoid theme due to a wide range of inference compute, control, sensing, power management, and charging requirements driving the semiconductor bill-of-materials assumptions in our Global Robot Model and supporting our humanoid semiconductor stock expressions.

With respect to economics, one humanoid leased at $5 an hour can replace two human workers making $25 an hour, supporting an NPV of approximately $200,000 per humanoid.

Humanoid robots seem to have captured the attention of private equity, with more than a doubling of global venture funding for the theme in 2025, following nearly a tripling of funding the year before.

The Morgan Stanley Global Robot Team is carefully tracking a range of humanoid robot efforts in both public and private markets globally.

On the public side, Morgan Stanley continues to curate and maintain the Humanoid 100, a global mapping of equities across a range of sectors and regions that may have an important role in bringing robots from the lab to your living room.

We divide the robot into the brain, body, and integrators: semiconductors, sensors, batteries, actuators, encoders, harmonic reducers, screws, bearings, connectors, and rare earths.

These include each of the Magnificent Seven, Palantir, Timken, Harmonic Drive, Hexagon, Regal Rexnord, Aptiv, Mobileye, NSK, SKF, STMicro, Samsung, Xiaomi, CATL, JL Mag, MP Materials, and many more.

We look forward to an exciting year of scientific innovation and capital formation as we follow the next chapters in the autonomous Industrial Revolution.

The robots are coming.

Stay human.