Researchers from top universities in the United States and China worked together last year teaching a dog how to do parkour. 

This pooch was no living and breathing canine, however. Instead it was an autonomous robot built using hardware from Unitree, a leading Chinese robotics firm. A team of seven scientists — from Stanford University and Carnegie Mellon University in the United States, and Tsinghua University, ShanghaiTech University, and the Shanghai Qi Zhi Institute in China — were able to program the robodog to recognize barriers and get around them, making it the first open-source project to achieve “athletic intelligence” without using real-world reference data, according to Stanford.

Global companies have been developing such robot dogs for years, marketing them as mobile security cameras, or as tools to help provide connectivity after major infrastructure damage.

Their most prominent potential use lies in military settings, however. Both the United States and China have touted robotic dogs’ potential to perform tasks like patrolling remote areas or disposing of explosive ordnances. Autonomous robots that do not require any real-world training data could be particularly effective tools for exploring unfamiliar territory behind enemy lines, experts say.

But while the parkour robodog project — which culminated in a peer-reviewed paper published last year — may have been successful scientifically, such examples of collaboration between American and Chinese scientists in areas with military applications raise questions about the trade-offs between the benefits of academic cooperation and the interests of national security. 

Researchers at U.S. universities must already comply with some government and university oversight in hi-tech fields, but these restrictions do not prohibit collaboration with Chinese peers. As a result, academic researchers in the United States and China are still working together in areas like robotics, as well as artificial intelligence and quantum computers — even as the U.S. government has stepped up its efforts to cut off Beijing’s access to similar technologies through export controls and outbound investment screening.

“You’ve got the Commerce Department doing one thing and researchers doing something else,” says William Hannas, lead analyst at Georgetown University’s Center for Security and Emerging Technology.

“The United States government can do a much better job in providing information to universities as to what the nature of the threat is,” he says, adding that he believes it lies in China’s systematic intellectual property theft and other “informal acquisition practices.”

The background to the Stanford-led robodog project demonstrates how complex Sino-American scientific collaboration can be, in part because the U.S. Department of Defense effectively has a dual role as a funder of science and a provider of the military. 

The U.S. Navy has been one of the project’s key backers, via the Office of Naval Research (ONR), which Congress established in 1946 “to plan, foster, and encourage scientific research in recognition of its paramount importance as related to the maintenance of future naval power, and the preservation of national security,” according to its website. In June 2020, ONR began funding Stanford computer scientist Chelsea Finn, a senior author of the final project paper, with a grant worth just over half a million dollars to explore how robots can “learn continuously from raw perception inputs” over a three-year period, according to the Defense Technical Information Center, an official repository of government-funded research. Finn did not respond to requests for comment.

In August 2020 — two months after Finn started receiving grant funding — ONR began prohibiting future recipients from procuring or obtaining video surveillance equipment from entities “owned or controlled by, or otherwise connected to” China’s government. According to WireScreen, Chinese state entities beneficially own 16 percent of Unitree, whose robots the Chinese military has used. The company is not on the entity list or otherwise sanctioned by the United States, but WireScreen shows that in 2022 it sold robots to the University of Electronic Science and Technology of China (UESTC), which the Department of Commerce added to the Entity List in 2012.

…would it be helpful if military robots were agile? Of course it would. But we’re just trying to make something that can catch up with a two-year-old.

Christopher Atkeson, a roboticist at Carnegie Mellon and author on the joint robodog paper

Unitree said in a statement that its products “are manufactured for civilian consumption” and that it is “not involved in deploying our products for any military purposes.” The company said it has no business with military-affiliated parties, and that customers may obtain its products through “third parties or other loopholes.” It said it has no records for a sale to UESTC. 

ONR said it had no direct funding connection to Unitree or knowledge of its use in the robodog project.

It is unclear whether ONR’s requirements now restrict federally funded researchers from using Unitree equipment in similar projects. Many robotics researchers in the United States use Unitree because of its low cost: A standard model costs $2,700 on the company’s website, while one from Boston Dynamics, a major American producer, costs around $75,000.

It is not known how much the local government-run Shanghai Qi Zhi and the participating Chinese universities invested in the project. Of the final paper’s seven authors, three are based at Chinese institutions, including Hang Zhao, a professor at Tsinghua who this year led one of 10 projects that won China’s “highest level official award in the field of artificial intelligence,” according to Shanghai Qi Zhi, where he is also a researcher.

Stanford, Carnegie Mellon, and Shanghai Qi Zhi did not respond to requests for comment.

Some scientists argue it is meaningless to talk about academic research being “stolen,” even if it has military applications. They note that their work is published widely and freely available, often with downloadable code. The robodog experiment’s code and methodology is accessible on Github, for example.

Christopher Atkeson, a roboticist at Carnegie Mellon who is an author on the joint robodog paper, says the project was about “skill and agility,” not creating military hardware.

“It’s like having somebody be a good football player. That’s not something that immediately leads to being a super soldier,” says Atkeson, whose previous work inspired the Disney movie Big Hero 6. “Now, would it be helpful if military robots were agile? Of course it would. But we’re just trying to make something that can catch up with a two-year-old.” 

Academics more broadly often argue that international collaboration is a boon to science. Research by Caroline Wagner, a professor at Ohio State University who studies the relationship between science and policy, shows that papers coauthored by U.S. and Chinese scientists across a range of disciplines were far more likely to be highly cited as those authored by scientists from just one of the countries.

But some critics say that the robodog should have raised alarm bells, and that even if other dual-use projects are themselves not problematic, universities should be more cautious about picking their partners.

“This is part of a larger, troubling trend,” says Craig Singleton, senior fellow for China at the nonpartisan Foundation for Defense of Democracies. “While not every collaboration is high-risk, there’s a need for transparent documentation of these partnerships and assurance that U.S. funds are not supporting research that could benefit China’s military ambitions.” 

In September, the House Select Committee on the Chinese Communist Party published a report identifying almost 9,000 academic publications supported financially by the Department of Defense and coauthored by researchers affiliated with Chinese institutions. Most of the publications were related to dual-use, critical and emerging technologies, according to the report. 

“The troubling conclusion then is that Department of Defense-funded research intended to allow the U.S. military to maintain a technological edge over its adversaries — has likely been used to enable and strengthen the PLA,” the report says.

“This is not surprising, because DOD policy permits it,” it adds.

Lawmakers are already stepping up their action in this area. In June, the Republican-led House passed the annual National Defense Authorization Act, including a clause that would cut off all Department of Defense funding for researchers who work with Chinese peers. The act has not yet passed the Senate, which did not include the clause in its draft version of the legislation.

An excerpt (right) from the National Defense Authorization Act, passed by the House on June 14, 2024, by a vote of 217-199 (left). Source: U.S. Congress, C-SPAN

Such efforts could intensify under President-elect Donald Trump’s new administration. During his first term in office, Trump launched the so-called China Initiative — later ended by President Joe Biden — that saw several academics with Chinese links charged with fraud and espionage. Many of the cases resulted in dismissal or acquittal. 

Since Trump began the initiative, nearly all leading research universities have beefed up their compliance protocols. But given the lack of federal standards in this area, procedures vary widely. Noting these discrepancies, the National Science Foundation this year created a $67 million program to manage research risks, known as SECURE.

While not every collaboration is high-risk, there’s a need for transparent documentation of these partnerships and assurance that U.S. funds are not supporting research that could benefit China’s military ambitions.

Craig Singleton, senior fellow for China at the nonpartisan Foundation for Defense of Democracies

The question of whether liability should fall on the university or the individual researcher if something goes wrong is “one of the topics of the hour,” says Wagner, who formerly advised the White House Office of Science and Technology Policy. “Many of the boundaries and barriers are unclear.”

Back in China, meanwhile, Hang Zhao, the award-winning AI researcher who was a senior author of the parkour robodog paper, unveiled a new study in June based on that project. He and his colleagues at ShanghaiTech, Shanghai Qi Zhi, and Tsinghua had built another robot that can autonomously climb and hurdle obstacles, also using Unitree hardware. 

This time, the robot no longer looks like a dog: It looks like a human.

---