The effort, launched two years ago as a way to quickly buy low-cost autonomous weapons to counter China’s growing military capabilities, is now being shifted to a new organization over concerns it isn’t moving fast enough, the people said.

The move reflects frustrations over setbacks in the program known as Replicator, a signature effort of the Biden administration’s Pentagon that aimed to deliver thousands of air-, land-, and sea-based AI systems by August 2025. Then-Deputy Defense Secretary Kathleen Hicks announced the program in 2023 with the promise of technology that would be “small, smart, cheap.”

While Hicks requested $1 billion over two years for Replicator, some lawmakers have called for billions of dollars more in spending, arguing the total needs to be far higher to ensure success.

Some Replicator systems have been unreliable, or were so expensive or slow to be manufactured they couldn’t be bought in the quantity needed, according to people familiar with the matter. The Pentagon has also struggled to find software that can successfully control large numbers of drones, made by different companies, working in coordination to find and potentially strike a target—a key to making the Replicator vision work.

The Pentagon leadership has shifted the Replicator work to a new division under Special Operations Command known as the Defense Autonomous Warfare Group, or DAWG, in the hopes of accelerating the program and focusing on the most appropriate weapons.

Those involved in Replicator offer different reasons for the delays, but say the effort has largely been a success. Some point to the military services, who pushed to buy systems that weren’t ready to be fielded, while others say the setbacks were just a normal part of any ambitious attempt to fast-track technology.

In an email, Hicks said Replicator was on track for success when she departed the Pentagon in late January, and had jump-started the process of buying autonomous systems for the military.

The purpose of Replicator is to prepare for a potential conflict with China in the Pacific. Beijing has rapidly expanded its arsenal of ships, aircraft and high-tech weapons in recent years, and U.S. officials believe Beijing may be ready to seize Taiwan—a key U.S. trading partner—as early as 2027.

A conflict over the island would bring technological and logistical challenges, requiring ocean vessels and aerial drones to cross long stretches and work autonomously, even if radio and GPS communications are jammed. The drones would allow the U.S. to spread out the battlefield, confuse the enemy, overwhelm defenses and attack targets without significant loss of life or expensive equipment, defense officials say.

DAWG now has less than two years to deliver the drones the Pentagon says it needs, according to the people familiar with the matter. The tight timeline reflects the urgency with which officials believe the U.S. must be prepared to fight a war in the Pacific.

Replicator is now being overseen by the vice commander of Special Operations Command, Lt. Gen. Frank Donovan, a defense official said. In August, as he was taking over the program, Donovan attended part of an event in California that was supposed to showcase some of the whiz-bang technology Replicator had acquired—but also highlighted that the systems weren’t ready for prime time, according to people who participated in the exercise.

An unmanned boat made by BlackSea Technologies experienced a rudder failure and drifted away. The launch of an aerial drone made by Anduril Industries, a venture-backed defense company, was delayed due to a potential problem with the launch tubes. And the software running on several boats misidentified or failed to identify objects as expected, the people said.

“There were very, very good things that happened from Replicator,” said Anduril’s founder Palmer Luckey. “Could it have been done better? Could it have been more clear about what exactly they were doing? Yes, of course. But big picture, I don’t think it was that bad.”

A spokeswoman for the Silicon Valley-based Defense Innovation Unit, which ran the Replicator program until August, didn’t respond to requests for comment. Some of the participants said that the exercise was a success, and that risk-taking and mishaps are a feature, not a bug.

Last month, the Defense Innovation Unit’s director, Doug Beck, a former combat veteran and Apple executive who was appointed during the Biden administration, resigned.

While the Defense Innovation Unit helped get commercial technologies into the military, it had to grapple with many of the same bureaucratic problems that have long existed in the Defense Department, according to people familiar with the organization. Uniformed officers who lacked technical expertise held sway over which drones would be bought in large numbers, the people said, and some platforms required extensive work to make them operate autonomously.

Of the dozen or so autonomous systems acquired for Replicator, three were unfinished or existed only as a concept at the time they were selected, according to people with knowledge of the matter. Among Replicator’s shortcomings, officials said, is that the Defense Innovation Unit was directed to buy drones that had older technology, and it didn’t rigorously test platforms and software before acquiring them, other people familiar with the matter said.

One such misstep was the purchase of hundreds of BlackSea’s unmanned boats, known as the GARC, or Global Autonomous Reconnaissance Craft, according to people familiar with the matter. The boats aren’t designed for complex, long-range missions in the Pacific, and Navy officers pushed for them without a clear understanding of their technical limitations, the people said. The military repeatedly changed software stacks and added complications, leading to mounting costs and unnecessary setbacks, they said.

The Navy didn’t respond to requests for comment.

One of the largest Replicator acquisitions was of the Switchblade 600 drone, which had struggled to perform in Ukraine. An analysis from an Army intelligence center suggested that Switchblade would be vulnerable in conditions where communications were jammed—a feature of modern conflicts, people familiar with the matter said.

AeroVironment, which produces the Switchblade, said it has made extensive improvements to the drone based on years of work in Ukraine, so it can now perform much better against electronic warfare. But the Army declined to buy the newer models for Replicator because it would have caused delays, the people said.

The Switchblade costs around $100,000—an order of magnitude more than the small drones the Ukrainians and Russians are using. AeroVironment said its aircraft’s capabilities far exceed that of the typical cheap drone used in Ukraine, and can take out huge air defense or missile launching systems, justifying the price tag.

Integrating the technology also proved challenging. During an exercise last year in the Pacific called Project Kahuna, drones from different manufacturers connected by Anduril’s software struggled at times to coordinate and perform tasks when out of sight from the operator, said people familiar with the exercise.

But those involved in the effort say Replicator, which was always intended to be transferred to the military, still notched significant achievements in two short years: It helped buy, test and advance new drone systems, pushed advancements in autonomous technology and shaved years off the traditional weapons-buying process. That approach is now being used for other Pentagon efforts, they say.

“We wanted to fill gaps and create a more competitive marketplace. Let’s scale what’s scalable, and then let’s find other technology that might be promising,” said Aditi Kumar, former principal deputy director of the innovation unit. “I think the transition to [Special Operations Command] is natural at this point.”