Anthropic presented the results of the second phase of the Project Fetch experiment, in which AI model Claude Opus 4.7 learned to work independently with a commercial four-legged robot. The self-assisted model performed a number of tasks in connection to sensors, programming and navigation, outperforming the company’s employees. In each task, Claude was at least 10 times faster, and in four tasks that both human teams completed, the model outperformed the team without AI 37 times, creating 10 times less code.

However, Claude was unable to successfully complete the final task of autonomously returning the ball to the starting point. The model correctly defined the position of the ball and positioned the robot for the push, but failed to control the precise movements in real time. Anthropic explained that the problem lies in the closed cycle of control, where the system must continuously evaluate the result of each action and adjust commands according to environmental changes. It is in such scenarios that people retain the advantage over large language models.

The company pointed out that the improvement of Claude’s robotic capabilities was a side effect of the overall scale-up of models. Anthropic believes that the industry is entering an early era of "physical agent AI," where models are gradually moving from helping people to performing tasks in the physical world on their own. Previously, a similar transformation had occurred in programming and cybersecurity, where AI agents learned to work autonomously with software tools.