Anthropic's AI agent taught a robot dog to follow commands 20 times faster than humans
Anthropic has unveiled updated results from its Project Fetch experiment. This time, the Claude Opus 4.7 language model demonstrated an impressive leap in autonomy: it handled the setup and control of a four-legged robot 20 times faster than teams of human engineers working last year.
In August 2024, Anthropic employees with no robotics experience attempted to program a robot dog using AI. At that time, the model only helped humans find solutions faster. The new testing phase radically changed the approach: Claude Opus 4.7 worked almost entirely autonomously with minimal researcher oversight.
The neural network independently completed a full cycle of tasks:
- connected to video sensors and LiDAR;
- wrote a program for manual control;
- created a motion trajectory monitoring system;
- configured an object recognition algorithm.
Comparative analysis showed stunning results. The Opus 4.7 model proved to be 18 times faster than the team using older AI versions, and 37 times faster than humans working without chatbot assistance. Moreover, the neural network generated significantly more efficient code—its volume turned out to be 10 times smaller than that of human teams. This indicates a deep understanding of the task and an ability to optimize at a level beyond most engineers.
It is important to note that, according to the developers, this progress was a side effect of general scaling of language models. Anthropic did not introduce specialized algorithms for controlling hardware—the model simply learned to better understand the physical world through code.
However, not everything is perfect. Claude still struggles with precise physical actions in real time. The model managed to guide the robot to its target but failed at the task of gently pushing a ball to a specific point. This requires complex feedback, where humans still hold an advantage.
My expert assessment: We are witnessing the dawn of the era of "physical AI agents." Anthropic is right—in the coming years, neural networks will begin using standard tools and equipment as effectively as they currently work with code. But the weak point—interaction with an unpredictable physical environment—will remain the main challenge. The industry will have to solve the problem of adapting AI to the real world, where a millimeter error could cost the entire experiment.