- Claude assisted in the programming and operation of a Unitree Go2, automating much of the work at Project Fetch.
- The AI-powered team solved some tasks faster, such as walking and locating a ball, than the unaided group could.
- The interaction analysis revealed less confusion with Claude, thanks to easier connection and a more usable interface.
- The progress highlights both opportunities and risks: protocols and physical safeguards need to be strengthened when bringing LLM into the real world.
The new test of anthropic It focuses on an issue that is no longer science fiction: What happens when a language model coordinates a robot?. In Project FetchTheir Claude system helped operate a robot dog, with the aim of testing how far the robot could go. Physical AI moving from text to movement.
Beyond the headline, the experiment provides clear clues about capabilities and limitations: Claude automated much of the necessary programming so that the quadruped could perform physical actions, and It served as a catalyst for a team of people to advance more quickly in certain tasks.
AI and the physical world: from the laboratory to action
Anthropic, founded by former OpenAI researchers, has long studied the risks and practical applications of advanced models. This time, the hypothesis was straightforward: if an LLM increasingly masters coding and interaction with with , can begin to influence real objectsThe internal security team (red team) wanted to observe this transition in a controlled environment.
Researchers point out that current models do not yet fully govern a complex robot, but They anticipate that future versions will have more room for maneuver.Therefore, it is useful to analyze how humans rely on AI to program and orchestrate physical behaviors, especially in Humanoid robotsbefore that moment arrives.
How Project Fetch was designed
The challenge pitted two teams with no prior robotics experience against each other: one assisted by Claude and the other that programmed without AI assistance. Both teams had to take control of a Unitree Go2 robot dog using a remote control and write code, working with controllers and platforms such as Arduino Uno Q and also to perform tasks of increasing difficulty, from walking towards a point to locating an object.
The group with Claude was able to achieve some objectives more quickly, including the quadruped I would walk and find a beach ballThis was something the human-only team couldn't achieve under the test conditions. The key wasn't magic; the model generated and refined code, speeding up the connection with the robot and reducing friction.
Anthropic recorded and analyzed the work dynamics. In the transcripts, the team without AI expressed more frustration and doubt, while Claude's assistance It seemed to facilitate a more understandable control interface. and a smoother start-up. Even so, not all goals were met and autonomy was limited.
The chosen robot dog: Unitree Go2 and its purpose
The Go2 model, manufactured by Unitree in Hangzhou, China, was chosen for the evaluation. It costs around $16.900, a relatively tight figure compared to other equipment in the sector, and is used in remote inspection tasks, security patrols or tours in construction and manufacturing.
This quadruped can move independently, but in practice it depends on high-level orders or the control of a personAccording to recent market analysis, Unitree systems are among the most widespread, making them an attractive testing ground to see how far AI-assisted programming can push the boundaries.
What do the results reveal about LLMs?
The great language models no longer just write texts: in recent years they have specialized in generate code and manage with In Project Fetch, that ability translated into less time spent on repetitive programming tasks and a step-by-step guide to iterating over errors and adapting robot behaviors.
The prudent interpretation is that, although we are not talking about total control, AI lowers the barrier to entry for non-expert teams They enable a physical platform to perform useful actions. It's a qualitative change: from being mere text generators, LLMs are beginning to act as systems orchestrators.
Risks and safeguards: how to avoid scares
Giving AI the ability to act on machines introduces obvious risks: code errors, faulty data, or deliberate misuse These failures can have physical consequences. Industrial robotics learned long ago to mitigate these failures with independent protections. with .
In this context, experts suggest combining several layers: operational boundaries, auditing of generated code, and, above all, mechanical emergency switches and protocols that do not depend on the model. The Anthropic study is framed precisely within that preventive logic.
Emerging applications and necessary precautions
With the appropriate safeguards, the same approach could be applied to logistics, maintenance, inspection, or assistance in environments where human presence is complexThe idea is not to replace technicians, but to provide tools that accelerate configurations and allow for more adaptive responses.
For these benefits to materialize, it will be necessary to agree on safe practices, clear documentation, and responsible deployment criteriaOtherwise, technical advances may clash with public trust or with perfectly avoidable operational risks.
The Project Fetch experience suggests a turning point: Claude demonstrated that an LLM can shorten the distance between code and actionStreamlining real-world tasks in a quadruped robot, while reminding us that the leap into the physical world requires controls, rigorous testing, and a safety culture to match.
I am a technology enthusiast who has turned his "geek" interests into a profession. I have spent more than 10 years of my life using cutting-edge technology and tinkering with all kinds of programs out of pure curiosity. Now I have specialized in computer technology and video games. This is because for more than 5 years I have been writing for various websites on technology and video games, creating articles that seek to give you the information you need in a language that is understandable to everyone.
If you have any questions, my knowledge ranges from everything related to the Windows operating system as well as Android for mobile phones. And my commitment is to you, I am always willing to spend a few minutes and help you resolve any questions you may have in this internet world.