AI agents take on "hardware": Claude Opus 4.7 controls a robot dog 20 times faster than humans
The artificial intelligence market continues to surprise with its pace of evolution. A new experiment by the company Anthropic, known as Project Fetch, showed that language models are already capable not just of generating text, but of fully controlling physical devices. During the second testing phase, the Claude Opus 4.7 model independently programmed and configured a four-legged robot, completing the task 20 times faster than a team of human engineers did a year earlier.
For comparison: in August 2024, Anthropic employees with no robotics experience tried to make a robot move using AI prompts. At that time, the neural network only acted as an assistant. Today's version of Claude Opus 4.7 worked almost autonomously, under minimal researcher supervision.
What did the AI do?
The model performed a full cycle of hardware setup without human intervention. Specifically, Claude independently:
- connected to video sensors and LiDAR;
- wrote a program for manual robot control;
- created a trajectory monitoring system;
- configured an object recognition algorithm.
Notably, the code written by the neural network turned out to be 10 times more compact than that of human teams. This speaks not only to speed but also to execution quality. In numbers: Opus 4.7 was 18 times faster than a team using older AI versions and 37 times faster than humans working without chatbot assistance.
The paradox of physical interaction
Despite impressive progress in programming and logistics, Claude still faces serious difficulties with precise physical actions. The model successfully guided the robot to the goal but could not gently push a ball to the right spot. As the developers noted, the task requires complex real-time feedback—an area where humans still maintain an advantage.
It is important to emphasize that Anthropic did not implement specialized algorithms for hardware control. All progress is a side effect of the general scaling of language models. This confirms the hypothesis that universal AI agents can adapt to physical tasks without specialized training.
Analyst's perspective
We are entering the era of "physical AI agents," as Anthropic rightly notes. Already, neural networks can work with tools and equipment as effectively as with software code. However, in my view, the key barrier remains not computational speed, but tactile feedback and an understanding of real-world physics. Once AI learns to sense the environment as well as it senses context, the boundaries between the digital and physical worlds will finally blur.