Claude AI agent surpasses humans in controlling robot dogs: a breakthrough in physical robotics
Anthropic has presented the results of the second phase of the Project Fetch experiment, and they are impressive. The Claude Opus 4.5 model demonstrated the ability to fully autonomously program and control a four-legged robot — and did so 20 times faster than the best human team of engineers working with previous versions of the AI.
Recall that in August 2024, company employees with no experience in robotics attempted to program the same robot, and at that time the AI only acted as an assistant. Today, the situation has changed dramatically.
Full Autonomy: What Claude Did
In the new testing phase, Claude Opus 4.5 worked with virtually no human involvement. Under minimal researcher supervision, the neural network independently completed a full cycle of tasks:
- connected to video sensors and LiDAR;
- wrote a complete program for manual robot control;
- created a trajectory monitoring system;
- configured an object recognition algorithm.
The key metric is speed. Opus 4.5 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 code generated by the neural network turned out to be 10 times more compact than human-written code — indicating a fundamentally different level of optimization.
A Side Effect of Scaling
An important nuance: Anthropic did not implement specialized algorithms for controlling hardware. According to the developers, the progress in robotics became a side effect of the general scaling of language models. In other words, Claude learned to control the robot simply because it became smarter overall.
However, not everything is smooth. Despite success in navigation, the model failed at the task of carefully pushing a ball to a specific point. Precise physical actions require complex real-time feedback — here, humans still maintain an advantage.
Looking to the Future
Anthropic believes the industry is entering an era of "physical AI agents." If today neural networks work with code, tomorrow they will be able to control standard equipment and tools with the same efficiency.
My expert commentary: This experiment is an important signal for the market. We are seeing AI transition from a purely digital environment into the physical world. For crypto infrastructure, this means potential automation of mining farms, logistics, and even physical security. However, the problem of precise motor skills remains a bottleneck — and for now, it is precisely what separates us from fully autonomous robotic systems.