Anthropic's Claude Robot Dog Outcodes Humans but Fails at Fetch

In the second phase of Project Fetch, Anthropic pushed its AI to the limit by having Claude Opus 4.7 independently write the software stack for a quadrupedal robot. The model successfully generated the entire control system autonomously and finished the tasks roughly 20 times faster than the human teams, managing the workload in under ten minutes. While the code was concise, the robot still completely ignored the beach ball it was supposed to retrieve.

Comparing this to the first phase of the experiment, the progress is jarring. A year ago, the AI struggled to even establish a basic connection to the hardware. Now, it is producing nearly ten times less code than humans to achieve the same functional results, proving that the model's ability to navigate physical systems is scaling much faster than the robotic dexterity required to interact with a simple sphere.

The current wall is sensory feedback and precision. Claude lacks the real-time, closed-loop motor control necessary to track a moving object and adjust its trajectory on the fly. Interestingly, only one human researcher with specialized robotics experience managed to pull off the task, highlighting that this remains a gap even for most biological operators. The AI isn't failing because it's dumb; it's failing because the physical world doesn't play by the rules of clean, logical syntax.

This experiment signals a transition from AI as a fancy chatbot to AI as a physical agent. The industry is currently witnessing the same trajectory that once occurred with text-based tools, suggesting that it is only a matter of time before the hardware catches up to the software's ambition. Whether this leads to a helpful assistant or a very fast, very confused metallic pet remains the million-dollar question.

Source: Anthropic Red Team