Anthropic believes AI models will increasingly reach into the physical world. To understand where things are headed, it asked Claude to program a quadruped.
Anthropic asked its model Claude to program a quadruped robot dog (the Unitree Go2) and found that Claude could automate much of the work—even tasks that the human-only team couldn’t complete.
The experiment, called Project Fetch, compared two groups: one programming without assistance and one using Claude. The assisted team completed tasks such as having the robot walk and find a beach ball faster than the non-assisted team.
Anthropic frames this as evidence that AI models are beginning to bridge from software into the physical world, though researchers caution today’s systems are still limited and rely on external sensing/robotics interfaces.
Notably, this raises both new possibilities and new risks: as models gain agency in the real world, oversight and safety become more pressing.
