A new way to train factory robots could revolutionize how militaries make drones and other weapons, enabling high-volume manufacturing close to front lines. And beyond the battlefield, it shows a possible path forward in the next era of manufacturing, a topic central to the competition between the United States and China, which the National Defense Strategy calls the “pacing challenge.”
The paper, published in the January issue of the International Journal of Extreme Manufacturing, lays out a transformative vision of AI-driven additive manufacturing, or AAM.
Today’s factory robots only are capable of a set number of rigid movements, are difficult to adapt to new tasks, and require highly specialized places on production floors. They can’t see when they are lined up incorrectly, or when they make mistakes.
The new system—developed by an international team of researchers from California State University, Northridge; the National University of Singapore; NASA’s Jet Propulsion Laboratory; and the University of Wisconsin-Madison—uses highly-skilled engineers to train robots in a much fuller range of human-like movements, enabling them to perceive and (on a basic level) understand what they are doing. When combined with 3D printing technology, the framework opens up the possibility of complete end-to-end manufacturing of electronics, like the small drones re-shaping the battlefield in Ukraine.
But it also addresses a key question at the root of the overarching power struggle between the United States and China: how to increase the manufacturing strength of the U.S.
Before President Donald Trump made re-shoring manufacturing a core focus of his presidency, the Biden White House undertook a series of different policy moves with the same goal—including the use of smaller and more gradual tariffs accompanied by grants and other incentives.
The Pentagon has launched several initiatives to reshore production of things like microelectronics, citing China’s dominance in those sectors as a strategic liability. The CHIPS and Science Act of 2022, which allocates over $52 billion for domestic semiconductor manufacturing, reflects bipartisan recognition that industrial capability underpins military and economic strength. Without secure, domestic access to critical technologies, the U.S. risks operational delays, compromised weapons systems, and diminished deterrence capacity in future conflicts.
But the United States suffers from a critical lack of skilled workers to meet those goals. And institutes like NIST, as well as industry heads, forecast that even with government investment, building that workforce would take at least seven years.
Commerce Secretary Howard Lutnik said this month that robotics and automation could play a key role in a U.S.-led manufacturing renaissance. And several American robotics companies are pushing for a national strategy to promote manufacturing automation.
However, the current state of factory robotics lags behind hopes for a quick transition to factory automation. A 2019 report from Boston Consulting Group points out the obvious: manufacturing robots lack the adaptiveness and dexterity of human workers, and robots lack the problem-solving abilities that are still key to actually producing things. That’s changed little since 2019, as revealed by the controversies and embarrassing public demonstrations of the Optimus robot developed by Tesla. Questions about whether Optimus is indeed autonomous or just an elaborate human-controlled puppet point to larger uncertainties about exactly how to make truly humanistic robots.
That’s why this new paper is so significant. Traditional manufacturing methods, even robots, “need a lot of well-trained engineers,” said co-author Bingbing Li of California State University, Northridge. And even the most advanced 3D printers often require manual steps for design input, process selection, or post-processing.
The paper’s AAM vision includes end-to-end autonomy: from computer-assisted design model preparation and nesting to scheduling, process optimization, and post-print machining—all managed by a fleet of collaborating robots or drones.
This integration will be particularly important for manufacturing in outer space, hence the interest and support from NASA. Li said he’s also been in contact with military contractors and the Defense Department’s office of Operational Testing and Evaluation for a defense-related project that could begin in May, if final approvals are granted.
AAM relies heavily on what the authors call sensor-integrated design: the use of many sensors of visible light, heat, and other phenomena to create a sort of perception in a software “brain.”
That brain, which differentiates tomorrow’s factory robots from today’s, will have four layers: a “knowledge layer” at the base of the model collects data from sensors, simulations, and past operations. A “generative solution layer” uses AI tools like large language models and knowledge graphs to model decision-making. An “operational layer” implements decisions on hardware, software, and robotic systems; and a “cognitive layer” empowers machines with agency, allowing them to reason, act, and even reflect.
“Within this [cognitive] layer, AI agents act as high-level controllers, assessing the skill pool in the operational layer to select appropriate skills for task execution based on the current state,” the authors write. “These agents are responsible for planning and executing optimal actions, engaging in a lifelong learning process that enhances their expertise through continuous reflection and learning.”
That process of learning is similar to the way humans interact with large language models. At first, the results are generic and error-prone, but over time, the relationship between the trainer and the machine produces a type of human-like intelligence that can be replicated across multiple robots.
“We hope to minimize the human operation down to or close to zero, because the whole process itself, we want to automate,” Li said.
The degree to which some manufacturing processes can be automated will depend greatly on the complexity of the thing that is being produced and how quickly AI systems can learn to adapt to new, unexpected events. Li said human experts will still play a crucial role in that process, but more akin to supervisors than operators. Instead of doing the manual tasks beyond a robot’s ken, the human will help the robots develop, improve, and validate ways to deal with the unexpected.
Ideally, the AI would become smart enough to actually consider how to improve on human designs, via what the authors call “generative design tools,” including diffusion models, 3D reconstruction, and neural radiance fields. NeRFs extrapolate depth and volume from pictures, the same way the brain does, to automate the design and creation of 3D objects from mere sketches or photos.
“These advancements pave the way for 3D content creation, potentially bringing us closer to the realization of ‘What you think is what you get,’” the authors write. “By exploring vast design spaces, these tools generate innovative solutions that traditional methods might overlook.”
In a military context, this means a soldier in the field could scan a broken part and an AI-enhanced system could redesign and print a replacement with minimal oversight—a power that could vastly reduce manning and logistics needs at the front.
The Defense Department, including DARPA, have increasingly emphasized secure, scalable, and on-demand manufacturing for contested logistics environments. The AAM method of closed-loop manufacturing using multiple sensors, feeding data to a layered intelligence, offering a roadmap for not only smarter production but resilient, adaptive manufacturing capabilities that could redefine the industrial front lines.
But beyond bases, the framework shows a path forward for moving manufacturing back to the United States. That has implications for the broader race with China, which wields its industrial capacity not only for economic dominance but to produce vast weapons stockpiles.
Then-Undersecretary of Defense for Acquisition and Sustainment Bill LaPlante noted this challenge in October 2023: “Today, the U.S. is in a technological and economic race to maintain its manufacturing edge, particularly as it concerns critical defense systems, such as satellites, advanced munitions and communications technologies.”
But the advancement also highlights other serious challenges standing in the way of a more automated manufacturing future, namely that China is the market-share leader in key components that robots need, such as sensors and actuators. Additionally, China has much more access to the materials that would be required to print and build products—including weapons—in robot-run factories.
Robots will enable the United States to build more things. But they won’t build themselves.
The post Could this breakthrough in robot manufacturing reshape warfare? appeared first on Defense One.
