Claude Opus 4.7, operating entirely without human assistance, completed sensor connection, robot control, and object detection tasks roughly 20 times faster than the fastest human teams from a year ago, according to a new phase of Anthropic’s Project Fetch published June 18.
The experiment marks a significant milestone in physical agentic AI: an LLM autonomously writing control programs, interfacing with lidar and video sensors, and operating a physical robot with no human intervention beyond plugging in the laptop and approving commands.
In the original Project Fetch experiment from August 2025, human teams, both with and without Claude assistance, operated a Unitree Go2 robot dog to locate and retrieve a beach ball. The Claude-assisted team substantially outperformed the unassisted team, but a human remained in the loop at every step.
Phase Two, published this week, removed the human from the loop entirely. Claude Opus 4.7, running through Claude Code with adaptive thinking and maximum effort settings, was given a simple prompt: connect to the robot’s sensors, write a control program, detect the beach ball, and retrieve it. The model connected to video and lidar sensors, wrote its own navigation code, monitored the robot’s path through camera feeds, and successfully detected the ball.
On four tasks completed by all groups, Opus 4.7 averaged 9 minutes and 35 seconds. The fastest human team required 181 minutes. The unassisted team took 361 minutes. That is a 19-fold speedup over humans using Claude pair-programming, and a 38-fold speedup over humans working without AI assistance.
Where it works and where it doesn’t
Claude’s code was often effective on the first try, something no human team achieved. It wrote roughly 1,000 lines of code compared to the assisted team’s 10,000, yet accomplished more. The model quickly identified the best sensor interface approach, where human teams had struggled with multiple options.
The fetch itself, physically nudging the ball back to base, remains unsolved. The robot could position itself behind the ball but lacked the fine motor control to push it precisely across the finish line. Humans succeeded at the retrieval through iterative practice and physical intuition.
“This is a kind of closed loop at which people excel (at least after making some mistakes and learning from them),” the researchers noted.
Implications
Anthropic’s researchers point out that the improvements did not come from focused robotics research. They emerged from the general scaling of LLMs. “Models building their own software tools might have seemed outlandish not long ago, but it is happening,” the paper states. “It would be unwise to rule out the same trajectory in hardware.”
Sources: Project Fetch: Phase Two (Anthropic Research, June 18, 2026); Michael Ilie, C. Daniel Freeman, Kevin K. Troy.