NewsBB-0090

Project Fetch phase two: Claude Opus 4.7 in Claude Code writes the sensor code for a robot dog

Anthropic repeated its robot-dog experiment with Claude Opus 4.7 in Claude Code, a researcher approving each command. On every task a 2025 human team had completed, including connecting to the camera and lidar, Claude was at least ten times faster. Anthropic says its attempts to knock the beach ball back were poorly controlled, and it never finished the autonomous fetch.

Open original source anthropic.com · Anthropic Frontier Red Team

Claimed
Why this evidence label?

Charts and timing tables only; no video of the robot.

How we check entries
Preview image for Project Fetch phase two: Claude Opus 4.7 in Claude Code writes the sensor code for a robot dog
The source's own preview image, from anthropic.com.

Briefing

The 2025 task list was: drive the robot with its controller, connect to video and lidar, write a manual-control program, track its path, detect the beach ball and fetch it autonomously. The controller step was dropped, since Claude could not use a physical controller. Three trials ran Opus 4.7 at maximum effort. On the four tasks both human teams finished, Anthropic reports it was more than 37 times faster than the team without Claude and more than 18 times faster than Team Claude, with almost ten times less code than Team Claude.

Claude defaulted to an outdated object-detection algorithm, which slowed one ball-detection trial. A researcher with more robotics experience wrote the autonomous fetch program. A footnote says Claude Mythos Preview was left out because its trials were not comparable.

Tags

Report a problem with this entry