Claude AI agent surpasses humans in controlling a robot dog: analysis of Anthropic's breakthrough
Anthropic has presented the results of the second phase of the Project Fetch experiment, and the results are impressive. The Claude Opus 4.5 model demonstrated the ability to autonomously program and control a four-legged robot, completing tasks 20 times faster than the best teams of human engineers a year earlier. This is not just another benchmark—it is a signal of a paradigm shift in robotics.
In the first phase of the experiment, launched in August 2024, the AI only acted as an assistant for people with no experience in robotics. Now, Claude Opus 4.5 worked almost autonomously, under minimal researcher supervision. The neural network independently completed a full cycle of tasks:
- Connected to video sensors and lidar;
- Wrote a program for manual control;
- Created a robot trajectory monitoring system;
- Configured an object recognition algorithm.
The key metric is speed. Opus 4.5 turned out to be 18 times faster than a team using previous AI versions, and 37 times faster than people working without chatbot assistance. At the same time, the volume of written code was 10 times less than that of human teams. This speaks not only to productivity but also to the quality of solutions: the model generates more efficient and concise code.
It is important to emphasize that Anthropic did not implement specialized algorithms for controlling the hardware. According to the developers, the progress in robotics became a side effect of the general scaling of language models. This confirms the hypothesis that universal AI agents can adapt to physical tasks without additional training.
However, there are limitations. Claude still struggles with precise physical manipulations. The model managed to guide the robot to the target but failed at the task of gently pushing a ball. This requires complex real-time feedback, in which humans still outperform AI. Nevertheless, Anthropic is confident: the industry is entering an era of "physical AI agents," where neural networks will work with standard tools as effectively as with code.
My professional opinion: This experiment is an important step toward autonomous robots controlled by AI. However, the results should not be overestimated. So far, we are talking about controlled conditions, while the real world is full of unpredictable variables. Still, the speed of progress is impressive, and if Anthropic continues at the same pace, we will see practical applications of these technologies within the next 2-3 years.