For an industry that’s only just getting started, there’s a lot of hype around humanoid robots.

You can thank Elon Musk and his bold pronouncements around Tesla’s Optimus robot. Morgan Stanley also threw gasoline on the hype fire with its forecast of nearly 1 billion humanoids in service by 2050 amid a $5 trillion market. And then there are the slick internet videos showing human-shaped robots doing back flips, cartwheels and other spectacular feats.

The reality is that most people overestimate what robots can do at this point in their development. It’s also true that humanoid robots have advanced rapidly in the last decade when they were expensive laboratory experiments at universities and specialty firms such as Boston Dynamics. Now these mobile machines with arms are being deployed mostly as pilot projects in warehouses, factories and even hospitals.

With these early humanoid robot deployments, the barriers around safety, power supply and machine learning are much clearer now for achieving the goal of building general-purpose mobile robots that can perform multiple tasks around humans at a reasonable cost. These robots, which include models that have wheels instead of legs, represent the last evolutionary step for robots that began more than six decades ago as clunky, pneumatically driven machines that were bolted to the floor and caged off from workers.

The importance of developing a U.S. home-grown humanoid robot industry that includes assembly and the full supply chain — motors, actuators, sensors, chips, cameras, battery packs, etc. — can’t be overstated. Similar to shipbuilding and even drones, the products will be integral to the civilian economy with crossover uses for defense. General-purpose robots will allow economies to grow even as the human population hits a plateau and begins to decline. This native-born population decline, of course, is happening now in developed nations.

The skeptics focus on the limitations of these new machines, which now perform simple tasks such as fetching items, picking them from totes or placing plastic crates on conveyor belts. They are often slow, chew through battery power and take a long time to train. History tells us that technological improvements will overcome these barriers. History also points to the adoption taking more time than optimists predict.

To give a sense of how new this industry is, the Association for Advancing Automation held its second annual conference dedicated to humanoid robots in Seattle late last year. The crowd was bigger than the one at the first gathering, with many company engineers turning out to see whether these machines will work in their factories or job sites.

The answer: not yet. A dose of optimism in the morning from robot startups was countered by skepticism in the afternoon, including from Brad Porter, who led Amazon.com’s robotics organization before founding his own company, Cobot. Amazon’s analysis of industrial robots found only 40 use cases for humanoids that couldn’t be done by other types of robots, Porter said.

One of those uses is for a wheeled robot to haul around box-laden carts autonomously for the shipping giant AP Moller-Maersk A/S. Porter’s startup designed a robot, Proxie, for that job. Although the use cases are limited now, more will arise as the robots improve performance. Agility Robotics has deployed its robot, Digit, to several warehouses where it can pick up totes and walk over to place them on a conveyor belt.

The areas where the robots work, though, are closed to humans. There are safety concerns around legged robots, mostly because they crumple to the ground when the power is cut. Bipedal robots use power for "dynamic stability,” which keeps them balanced on their feet when standing or moving around. Battery power is a significant limitation for humanoid robots and the need for power will only increase as they become more capable of performing additional tasks.

The robots must host AI computing power because the delayed signal from cloud servers is too great, Amit Goel, head of Robotics Edge Computing Ecosystem at Nvidia, said during the one-day session.

One of the solutions is to tether the robots to a power supply. This works only if they don’t need to move around much. Another method is to have the robots swap batteries when the power drains to a certain level. This requires additional investment for extra power packs, space where the swap is made and time to change out batteries. A design solution is to put the robot on wheels instead of legs, eliminating the power demand for stability. The trade-off is that the wheelbase must become increasingly large if the robot is going to lift heavy objects with its arms. It’s also much harder to navigate a human world on wheels.

Wheels are the mode of locomotion for Moxi, the robot made by Diligent Robotics that can take medical supplies, medicines, laboratory samples and other objects directly to nurses. The robot can navigate the hospital by itself — taking the elevator, opening doors that require a badge or button to operate and maneuvering around people and objects. About 100 Moxi robots are deployed in more than 25 hospitals. This saves time and walking for nurses, even if a hospital has a pneumatic tube delivery system.

Moxi, for now, has one use case: delivering stuff to hospital personnel. Humanoid proliferation won’t come anywhere near the Morgan Stanley predictions if the robots can’t do multiple tasks and learn them quickly.

This is where the techniques for training the robots have to improve with simulation, Goel said. There’s a lack of data to train robots and other so-called physical AI. It’s costly and time consuming to train the robot by performing tasks. If the robot fails, it could damage itself or objects around it. Generic foundational models that depict real world situations will help reduce the cost of simulation and speed training.

Jeff Cardenas, co-founder and CEO of robotics startup Apptronik, said he dreaded going to automation conferences where participants would chuckle at the idea of a general-purpose humanoid robot and he’d get a pat on the back for trying. The consensus has changed along with the tremendous progress made since a Defense Advanced Research Projects Agency (DARPA) challenge from 2012 to 2015 jump-started the industry.

Apptronik is working with Google DeepMind to address safety issues and has partnered with Jabil to both manufacture and use the startup’s Apollo robot. Cardenas said this year will still be about pilots and early orders, with the industry deploying hundreds of humanoid robots. "I think 2027 is the year when you start to see early scale.”

The era of humanoid robots is about to begin and the U.S. and other governments need to control over their own destinies.

© The Japan Times