Anthropic has updated its Project Fetch experiment. The Claude Opus 4.7 model completed tasks for setting up and controlling a robot dog 20 times faster than teams of human engineers.

In August 2024, company employees with no experience in robotics tried to program a four-legged robot. Back then, the AI only helped people find solutions faster. In the new testing phase, the Claude Opus 4.7 model worked almost autonomously under minimal supervision from a researcher. The neural network on its own:

• connected to the video and lidar sensors;
• wrote a program for manual control;
• created a system for monitoring the robot's path;
• configured an object recognition algorithm.

The Opus 4.7 model turned out to be 18 times faster than the team that used older AI versions, and 37 times faster than people without the help of a chatbot. The neural network wrote more efficient code: its volume was 10 times smaller than that of the human teams.

The authors noted that progress in robotics became a side effect of the general scaling of language models. Anthropic did not implement specialized algorithms for controlling "hardware."

Despite the success, Claude still struggles with precise physical actions. The model was able to guide the robot to its goal but failed at the task of accurately nudging a ball into the desired spot. This requires complex real-time feedback, in which humans still outperform AI.

Anthropic believes the industry is entering the era of "physical AI agents." In the future, neural networks will be able to use standard tools and equipment as effectively as they now work with program code.

As a reminder, on June 13 Anthropic halted access to the Fable 5 and Mythos 5 models. The reason was a US government directive as part of export controls.

Source: ForkLog