Anthropic has released results from Project Fetch: Phase Two, a research effort that tested whether Claude, their AI language model, can meaningfully assist engineers working on robotics tasks. The original experiment ran in August 2025, with findings published on Anthropic's research blog in June 2026.

Project Fetch: Phase Two builds on an earlier phase documented in November 2025. The central question is practical: can a large language model like Claude act as a useful collaborator for people building robots and other physically-embodied systems — work that sits at the intersection of software logic and real-world mechanical action?

The project's name is literal in an instructive way. A robot dog tasked with fetching a beach ball during the experiment failed to complete the retrieval, a detail Anthropic did not downplay. It is a candid result — manipulating soft, rolling objects in unstructured outdoor environments remains one of the harder unsolved problems in robotics, and a language model alone does not straightforwardly solve the challenge of sensors and robot actuators working together in the real world.

What Project Fetch actually tests is narrower and arguably more workable: whether Claude can support the human side of robotics work — helping engineers think through how to design a task, debug why a system is not working as expected, make sense of sensor data, or navigate the complexity of setting up and running a robotic system. This is AI as a tool that makes skilled experts faster and more capable, rather than as a replacement for the robot's own perception and control logic.

That framing matters. Much of the public discussion around AI and robotics conflates two different questions: what can a language model do when its output is fed directly into a robot's control system, and what can a language model do when a human engineer is in the loop — reading the output, thinking about it, and using it to guide their next decision. Project Fetch, at least in these phases, appears to focus on the second scenario. That is a more measured scope and arguably more immediately practical — when companies deploy AI in the real world, they have found that giving skilled workers better tools tends to produce faster and more reliable results than trying to fully automate complex physical work.

Anthropicran this experiment with its own staff, a choice that reflects how the company evaluates Claude's capabilities. Using internal employees provides a controlled group of motivated users who understand the technical domain, while also revealing the real friction points that polished demonstrations often hide.

Robotics is a useful testing ground for this reason. Robotics work requires many sequential steps, demands tight feedback between planning and doing, and involves debugging across hardware, firmware, and software simultaneously. If Claude can add real value in that context, the same underlying capability — staying organized across complex, multi-step technical work — would likely transfer into other engineering fields.

The beach ball failure deserves attention, but not as proof of broader failure. Reliable robot manipulation of soft, unpredictably rolling objects in open space is still a challenge even for robots built with specialized sensors. Expecting a robot dog, presumably using general-purpose movement software, to reliably fetch a beach ball says as much about the physical and control limitations of the hardware as it does about any AI support system above it.

What Phase Two contributes is a more thorough accounting of where Claude genuinely helped with these tasks and where it fell short — a more instructive result from a research perspective. The progression from Phase One (November 2025) to Phase Two (June 2026) suggests Anthropic is using an iterative approach: define the task, run the test, publish what actually happened, refine the scope, repeat.

For engineers in robotics and automation fields, the practical question is whether Claude-class models are ready to work alongside them in a real project, not just in a carefully controlled demonstration. Project Fetch is a structured attempt to answer that question under conditions that resemble actual work. The results, including where Claude fell short, are more reliable input for that judgment than any isolated benchmark test.