Former NASA robotics chief: America is building the wrong kind of robots — and China knows it

Looking beyond the robot legwork, China’s demonstration reveals a widening chasm between spectacle and strategy—a gap into which America, for all its robotics prowess, is in real danger of falling.

We build impressive robots. We don’t build real ones.

Walk into any major robotics demo in the U.S. and you’ll see fluid movements, precise manipulation, and maybe even back flips. The most advanced Boston Dynamics robots can pick up and carry large objects that could injure any worker. Based on performance alone, we look competitive.

The problem is that we only measure performance in controlled settings. According to a recent Stanford report, robots that achieve a success rate of nearly 90 percent in controlled simulations succeed in only 12 percent of real household tasks. This difference between demo and deployment isn’t a rounding error – it’s the whole problem. The US optimizes its humanoid robots for sprinting and calls it a marathon strategy.

Take the 02 model of Figure AI for example. It logged 1,250 hours at BMW’s Spartanburg plant and moved more than 90,000 components. By today’s standards, it was a success. Take a closer look, and the robot did one job: pick up sheet metal parts and place them on a welding stand for 10 months straight. A medium-sized manufacturer—already using automated systems—cannot justify investing thousands of dollars in a machine that does one thing.

Successful one-off robot deployments obscure the real question: is this investment worth it at scale?

What NASA taught me about fragile machines

At NASA, decades of designing humanoid robots for environments that don’t forgive narrow thinking revealed that the machines that failed were built for a single scenario. Those that succeeded could be multitasked and reprogrammed– for use in different settings. For example, an arm built for the space shuttle was designed to locate an astronaut who would catch and later release a satellite. It turned out that the robot was better at making the catch itself—but locating astronauts proved useful for other tasks, like repairing the Hubble Space Telescope.

The US continues to build humanoids that are exceptional in the conditions they are trained for and fragile everywhere else. Currently, in most factories, multiple humans produce a better ROI than a single humanoid robot.

We humans may not be the strongest or fastest, but we make up for it with adaptability. A single warehouse worker can pick orders, restock shelves, report a safety issue, and reroute around a spill—all before lunch. This smooth change of tasks is the core of a person’s work value. If humanoid robots are supposed to take over US factory floors, they need to be designed to be even more flexible than we are.

The political environment is not ready either

To get there, we must first prepare the political environment. American manufacturers—especially the midsize and enterprise manufacturers that form the backbone of US industrial production—have little to no structured way to deploy humanoid robotics at scale. Large manufacturers such as BMW may adopt a ten-month single-task pilot as an R&D line item. A medium-sized car supplier or contract manufacturer cannot.

The absence of a proper federal incentive structure prolongs stagnation. Current federal R&D tax credits reward robotics discovery, not deployment. A manufacturer who spends $800,000 to integrate a humanoid system gets essentially the same tax credit as a buyer of a new forklift.

Although $2.5 billion in venture capital has already been poured into robotics, private investment alone cannot spur effective adoption. Instead, the nation needs a separate one, a robotics-focused “manufacturing adoption” tax incentive that stacks with an existing R&D credit—rewarding those who make robots work in real factories by offsetting integration costs, labor transition costs, and process redesign efforts. The United States may also expand the Manufacturing Extension Partnership, which already provides specialized consulting to small and medium-sized manufacturers, to provide humanoid concierges at relatively low federal cost. Finally, NIST, working with NASA and others, should create humanoid interoperability standards to enable manufacturers to connect multiple robotic systems safely.

What a proper deployment actually looks like

U.S. factories must also change. Most industrial workflows were designed around human flexibility, improvisation and self-direction. Robots—even adaptive ones—require something different: fleet-based missions similar to ride-on routes, clear safety parameters for human-robot environments, and new protocols for humanoids to interact with single-use machines.

Both the US and China continue to misunderstand the assumption that humanoids will replace jobs wholesale. Rather, they create value by filling the gaps that current automation can’t reach—the « middle work » that’s too variable for a fixed conveyor system and too repetitive to justify a skilled worker. A medium-sized company appreciates being able to transfer humanoids to single-use machines, such as loading a washing machine, without having to change the washer itself.

Moving materials between workstations, adding inventory to busy warehouses, tending machines built for human interaction, and performing inspections in dangerous and confined spaces are not great use cases. But they represent functional problems that adaptive humanoids may begin to address within this decade.

America has the talent, capital and industrial base to lead this transformation. We are currently optimizing for false results and ignoring practices that can enable real deployment.

A country that defines « good enough to deploy at scale » sets the terms for global manufacturing for decades. Currently, that country is not the United States.

It doesn’t have to stay that way.

Opinions expressed in Fortune.com comments are solely those of their authors and do not necessarily reflect their opinions and beliefs. Luck.