Humanoids need to get better at retaining information and learning from their mistakes. Memory technology will be key to future breakthroughs.
Humanoid-robot maker Figure AI is one of the companies Bernstein analysts see leading the development of modern robotic brains.
The next "great breakthrough" on the road to creating useful humanoid robots will be making their brains capable of retaining information, according to analysts.
Researchers have been working toward that goal for a long time, though the artificial-intelligence boom has helped speed things up, experts say. One recent breakthrough, Bernstein analyst Jay Huang noted, is the ability for humanoids to simulate potential scenarios and plan for them through AI models. But Huang and others in the industry think that can still be improved.
"How wonderful it would be if a robot, failing a task, immediately adjusts its actions based on the specific factors behind that failure, so its actions quickly converge toward success," Huang wrote in a client note on Friday.
The bottleneck there is memory - not a lack of memory chips, like the kind made by Micron Technology (MU), but the ability to use and retain information. Huang noted that fixing that issue is "already a new frontier of research."
Improving robotic brains is one of the key ways that the recent wave of humanoid firms can make their robots actually useful. The brain, J.P. Morgan analysts said in a client note last month, is the "largest bottleneck" keeping the industry from scaling.
Morgan Stanley has modeled the humanoid market being worth $5 trillion by 2050. By that point, more than 1 billion robots could be deployed in a variety of situations, according to the firm. These robots should be able to do factory work, like sorting packages, and proponents also envision humanoids helping out with household chores and childcare, among other things.
Huang said that world models, which are designed to comprehend real-world physics, may still struggle to actually achieve that goal in three years. He's more positive about robotic brains becoming better at fusing multiple sources of data and turning that into action toward the end of this decade.
Bernstein named Physical Intelligence, a two-year-old startup based in San Francisco, as a likely leader in robotic-brain development. The private company has raised more than $1 billion and was valued at $5.6 billion last November, according to FactSet. Amazon.com (AMZN) founder Jeff Bezos and OpenAI are among its investors.
In March, Physical Intelligence developed what it called Multi-Scale Embodied Memory, which allows its model to keep track of tasks for up to 15 minutes and correct mistakes - a significant feat. A month later, it revealed a model called <PI>0.7, which was able to perform tasks that "were never seen in its training data."
Other firms that Bernstein thinks will pave the way forward on brain models include Nvidia (NVDA), Alphabet (GOOG) (GOOGL) and private humanoid firm Figure AI. Figure last month revealed plans to rapidly scale its real-world data collation and get access to up to 100,000 Nvidia graphic processing units to help it advance its model development. It was most recently valued at $39 billion.
"Humanoid robots are the ultimate deployment vector" for artificial general intelligence, Figure CEO Brett Adcock wrote on X recently. He added that Figure will have "unprecedented" data and AI computing power.
While Figure is private, investors can get exposure to the company through RoboStrategy $(BOT)$, a fund that has purchased either direct stakes or participated in special-purpose vehicles that have bought shares in private humanoid firms. Through an SPV, Figure is RoboStrategy's fourth-largest holding.
Ulrike Hoffmann-Burchard, global head of equities for UBS Wealth Management, also recommends that investors interested in humanoids take a look at the companies enabling them.
"Physical AI is here, and it is going to unlock economic value across logistics, manufacturing and healthcare," Hoffmann-Burchard said in a recent note. "We believe that the companies making the gears that move humanoid robots, the sensors that let them see and the chips that let them think will become more valuable over time."
J.P. Morgan analysts have suggested Timken $(TKR)$, Regal Rexnord (RRX) and Allient $(ALNT)$ as potential "picks and shovels" winners of the robot revolution. Each has reported contracts for components that help humanoid limbs function - a total addressable market that J.P. Morgan estimates could be worth $4 billion by 2030.
-William Gavin