Anthropic, a company that builds AI systems, recently published results from an experiment called Project Fetch. They asked a simple question: can their AI assistant Claude help people who are building and working with robots?

The experiment ran in August 2025, and the results came out in June 2026. Project Fetch: Phase Two builds on an earlier test from November 2025.

The key finding is this: Claude was not tested as the robot's brain. Instead, researchers wanted to know if Claude could help the human engineers — the people designing, fixing, and managing the robots — work faster and smarter.

To test this, they asked a robot dog to fetch a beach ball. It did not succeed. Anthropic was open about this failure. A robot picking up and carrying something soft and round in an outdoor space is actually quite hard, even for robots. A language model, by itself, cannot fix that problem.

So what did Claude do instead? It helped the engineers think through how to design tasks, understand why the robot was not doing what they wanted, make sense of the information coming from the robot's sensors, and manage all the complicated steps involved in running a robot. Think of it like this: Claude was not the robot's hands, but it was something like a smart notebook that could help the engineer figure things out.

This is an important distinction. When people talk about AI and robots, they often mix two different ideas: AI controlling the robot directly, and AI helping the human who controls the robot. Project Fetch was testing the second idea. In the real world, companies have found that giving skilled workers better tools tends to produce better results faster than trying to take humans completely out of the loop.

Anthropicran this test with its own employees, all of them trained engineers. This meant they could watch real engineers actually using Claude in their day-to-day work, not just in a polished demonstration. Real work always reveals problems that perfect demos hide.

Robotics work is hard because it involves many steps, constant back-and-forth checking whether things are working, and debugging across physical hardware, firmware (the software that runs hardware), and regular software all at the same time. If Claude could help in that complex environment, the same skill would likely help in other types of engineering work too.

The beach ball failure matters for context. A robot reliably picking up something soft and bouncy in an open space challenges even robots that were specifically designed for that task. Blaming the AI for that failure would be like blaming a calculator for not being able to carry heavy boxes. The robot's physical limitations played a big role.

What Anthropic added with Phase Two was a careful look at what Claude got right and where it fell short — the real learning. They ran Phase One, looked at what happened, published their findings honestly, made adjustments, then ran Phase Two. This step-by-step approach is more useful than running a single perfect test in a lab.

For engineers in robotics and related fields, the real question is simple: can Claude actually help them do their job better, in real projects, not just in a demonstration? Project Fetch is one attempt to answer that under conditions that match how people actually work. The failures matter as much as the successes.