Anthropic has demonstrated how AI controls a robot dog tens of times faster than humans.
Anthropic has presented the results of the second phase of the Project Fetch experiment, and the results are impressive. My team and I are closely monitoring this project, as it demonstrates how far language models have advanced in autonomously solving complex physical tasks.
The Claude Opus 4.7 model completed the full cycle of setting up and controlling a robot dog 20 times faster than teams of human engineers working with previous versions of AI. This is not just an improvement in metrics — it is a paradigm shift.
How It Works
Unlike the first phase, where AI acted only as an assistant for people without experience in robotics, the new version of Claude worked almost autonomously. Under minimal researcher supervision, the model independently:
- connected to video sensors and LiDAR;
- wrote a program for manual control;
- created a system for monitoring the robot's path;
- configured an object recognition algorithm.
What is especially important: the neural network wrote code that is 10 times smaller in volume than that of human teams. This means not only speed but also quality — more efficient and concise solutions.
The Numbers Speak for Themselves
Opus 4.7 turned out to be 18 times faster than a team using older versions of AI, and 37 times faster than people working without the help of a chatbot. At the same time, Anthropic did not introduce specialized algorithms for controlling the hardware — the progress was a side effect of the general scaling of language models.
Limitations and Prospects
However, there were also weak points. Claude successfully guided the robot to the target, but failed at the task of gently pushing a ball to a specific point. This requires complex real-time feedback, where humans still outperform AI. Nevertheless, I believe this is a temporary limitation — once models learn to work with tactile feedback, physical agents will become a reality.
Anthropic is confident that the industry is entering an era of "physical AI agents." My professional forecast: in the next 2-3 years, we will see neural networks begin to use standard tools and equipment as effectively as they currently work with software code. This will open new horizons for automation in logistics, construction, and even household chores.