Claude's AI operates a robot dog 20 times faster

Claude Opus 4.7 completed several tasks with a robot dog up to 20 times faster than the fastest human team, according to an Anthropic experiment. The model worked with almost no help: one person only had to connect it to the robot, enter the initial instruction and approve its actions.

The experiment is part of the second phase of Project Fetch, a test designed to measure how far AI can go with a commercial quadruped robot. The goal was not to stage a science-fiction demonstration, but to solve specific tasks: access the sensors, write control programs, detect a ball and move the robot through a space.

From helping with programming to working on its own

In the first phase, conducted in August 2024, two teams of Anthropic employees tried to operate the robot. One could use Claude, while the other had only the internet and its own knowledge. The team that worked with Claude completed more tasks and did so in less time.

At that point, the model could not complete the process on its own. It even got stuck on something as basic as connecting to the robot. The new test aimed to determine whether newer models had closed that gap.

The answer was yes, although with clear limits. On every task that at least one of the human teams completed, Opus 4.7 took at least ten times less time. Looking only at the four tasks completed by both teams, the model was, on average, more than 37 times faster than the team without Claude and more than 18 times faster than the team that used it.

Claude quickly identified how to access the cameras and the lidar sensor, a technology that measures distances using light. It also wrote code that worked correctly on the first attempt in many cases. To achieve similar results, it produced almost ten times less code than the team that worked with Claude.

The robot still cannot search for the ball properly

The hardest part was the one that gives the project its name: getting the robot to pick up the beach ball and carry it back to its starting point.

Claude managed to position itself behind the ball and push it in the right direction. But it could not control the force and movement precisely enough. When the ball veered off course, the model did not adjust its commands quickly or delicately enough to correct its path.

People are better at this kind of continuous control. They observe the error, connect that result to the previous command and change the next movement almost instantly. It is a basic skill for handling physical objects, but it remains difficult for a model that works mainly with instructions and code.

Anthropic clarifies that this test does not show that AI has solved robotics. The experiment did not evaluate low-level tasks, such as creating a specific movement policy to control each of the robot's motors. In addition, a person with more robotics experience did manage to program the robot to autonomously retrieve the ball.

What this changes for you

The result points to an important transition. Until now, models could help a person program a robot. Now they are beginning to complete much of the preliminary work on their own, such as connecting sensors, choosing a method and writing the necessary software.

That could lower the barrier to using commercial robots for limited tasks, such as inspecting a space, following a route or detecting objects. It does not mean you can put an AI in charge of any machine without supervision. In physical environments, an error can cause damage, and precision remains a problem.

Anthropic says these improvements do not come from robotics-specific training, but from the general progress of language models. The next step will be to see whether Claude can maintain the same speed and reliability when it has to control more delicate movements, design its own tools or adapt to unfamiliar robots.

The experiment suggests that AI is beginning to enter the era of physical agents: systems capable of using tools in the real world, not just programs. The gap between writing code for a robot and operating it autonomously is still large, but it is narrowing faster than many expected.