Anthropic has tested how far Claude can go when its coding abilities are pointed at a machine that moves through the real world. In a study called Project Fetch, researchers used the AI model to help program a Unitree Go2 quadruped, a robot dog used for tasks such as remote inspections and security patrols.
The result was not a full AI takeover of a robot. It was more practical, and in some ways more important: Claude helped automate parts of the programming process, made the robot easier to work with, and let one team complete certain physical tasks faster than a team writing code without AI assistance.
What Anthropic Tested
Project Fetch was designed around a direct comparison. Anthropic asked two groups of researchers with no previous robotics experience to operate and program the Unitree Go2 quadruped. Both groups had access to a controller and were asked to work through increasingly complex activities.
One group used Claude’s coding model. The other group wrote code without AI help. That setup made the experiment less about whether expert roboticists could use AI effectively, and more about whether an AI coding assistant could lower the barrier for people who were new to robotics.
The Claude-assisted group did not succeed at everything. But it did complete some tasks faster than the human-only group. One example stood out: the team using Claude got the robot to walk around and find a beach ball, while the group without AI assistance could not solve that task.
That matters because the experiment connects two areas that have usually been discussed separately. Large language models have become useful at writing code and operating software. Robots, by contrast, require software to affect physical space. Project Fetch sits at the point where those two trends begin to overlap.
Why A Coding Model Helped A Robot Move
The source of Claude’s advantage was not described as raw robotic intelligence. The article says Claude was able to automate much of the work involved in programming the robot and getting it to perform physical tasks. In practice, that means the AI model helped with the software layer that sits between a human instruction and a robot’s behavior.
Anthropic also studied how the two teams worked together. The researchers recorded and analyzed interactions from both groups. They found that the group without Claude showed more negative sentiment and confusion.
One possible reason is that Claude helped the team connect to the robot more quickly and created an easier-to-use interface. That point is important because the biggest near-term effect of AI in robotics may not be robots making independent decisions. It may be AI making difficult robotic systems more accessible to people who do not already know the tools.
Changliu Liu, a roboticist at Carnegie Mellon University, said the results were interesting but not hugely surprising. She also pointed to the team-dynamics analysis as notable, because it suggests new ways to think about interfaces for AI-assisted coding.
“What I would be most interested to see is a more detailed breakdown of how Claude contributed,” she adds. “For example, whether it was identifying correct algorithms, choosing API calls, or something else more substantive.”
The Robot In The Experiment
The robot used in the experiment was the Unitree Go2 quadruped. It costs $16,900, which the source describes as relatively cheap by robot standards. The Go2 is typically used in industries such as construction and manufacturing for remote inspections and security patrols.
The robot can walk autonomously, but it usually depends on high-level software commands or a human using a controller. That makes it a useful test case for Claude: the model did not need to invent robotics from scratch, but it did need to help researchers work through the software needed to control a physical system.
Go2 is made by Unitree, a company based in Hangzhou, China. The source also says Unitree’s AI systems are currently the most popular on the market, according to a recent report by SemiAnalysis.
From Software Agents To Physical Action
Large language models are often associated with chatbots that generate text or images after a prompt. More recently, they have become better at writing code and operating software. That shift is what turns them into agents: systems that can take steps inside digital environments rather than only produce responses.
Project Fetch points toward the next question. If AI agents can operate software, what happens when that software controls machines?
Logan Graham, a member of Anthropic’s red team, framed the concern around models moving beyond screens and documents. He told WIRED:
“We have the suspicion that the next step for AI models is to start reaching out into the world and affecting the world more broadly,” Logan Graham, a member of Anthropic’s red team, which studies models for potential risks, tells WIRED. “This will really require models to interface more with robots.”
Anthropic was founded in 2021 by former OpenAI staffers who believed AI could become problematic, and even dangerous, as it advances. Graham said today’s models are not smart enough to fully control a robot, but future models might be. He also said studying how people use LLMs to program robots could help the industry prepare for “models eventually self-embodying.”
The Safety Question
The article does not say that Claude acted with malicious intent, or that an AI model currently has a clear reason to take control of a robot. It says that remains unclear. But Anthropic’s work is partly about looking ahead to risky possibilities before systems become more capable.
George Pappas, a computer scientist at the University of Pennsylvania, warned that using AI with robots can increase the potential for misuse and mishap. He said:
“Project Fetch demonstrates that LLMs can now instruct robots on tasks,” says George Pappas, a computer scientist at the University of Pennsylvania who studies these risks.
Pappas also noted an important limitation. Today’s AI models need access to other programs for sensing and navigation before they can take physical action. His group created a system called RoboGuard that restricts how AI models can cause a robot to misbehave by applying specific behavioral rules.
For now, Project Fetch is not a story about an AI model independently taking over a machine. It is a story about AI making robot programming easier, faster in some cases, and potentially more powerful as models improve. That could make robots more useful. It also gives researchers a clearer reason to study how AI agents should be constrained before they become more deeply connected to the physical world.