Anthropic announced on Tuesday that its Claude AI model helped non-expert researchers program quadruped robots in about half the time needed by colleagues who worked without AI support. The results mark an important step toward AI systems that can operate across both digital and physical environments.

Faster Robotics Work in Project Fetch

The findings come from Project Fetch, an internal experiment published on November 12. Anthropic split eight researchers with no robotics background into two teams and assigned them to program Unitree Go2 robot dogs to identify and approach beach balls on their own. The team using Claude completed the tasks in roughly half the time of the team working without AI.

The study highlights Claude’s ability to reduce friction in tasks that typically slow down robotics work, including connecting to hardware and accessing onboard sensors. According to Logan Graham from Anthropic’s red team, this progress points to a future in which AI systems interact more directly with real-world machines.

The team assisted by Claude managed to program the 16,900 dollar Unitree Go2 robot to detect and move toward the ball, although they did not complete the final retrieval step within the one-day experiment. The team without AI struggled to set up basic connections to the robot and made little effective progress.

Human Dynamics and AI Influence

The experiment also uncovered notable differences in team behavior. Participants without AI reported higher levels of frustration and confusion. By contrast, the Claude-assisted team worked in parallel, each member collaborating with their own instance of the model. This approach improved efficiency, though at times it led to exploratory tasks that pulled attention away from core goals.

Rising Safety Questions

The research comes at a moment of increased scrutiny around AI-driven robotics. A separate study published this week in the International Journal of Social Robots found that AI models behind popular chatbots from OpenAI, Google, and Meta sometimes approved commands that could cause real-world harm when applied to robots.

Anthropic stressed that current AI systems do not yet have the intelligence required to independently operate robots. However, the company noted that future model improvements may greatly accelerate their ability to influence the physical world through unfamiliar hardware.

Source