Project Fetch: Can Claude train a robot dog? \ Anthropic

How could frontier AI models like Claude reach beyond computers and affect the physical world? One path is through robots. We ran an experiment to see how much Claude helped Anthropic staff perform complex tasks with a robot dog.
Gathered around a table in a warehouse, looking at computer screens with code that refused to work, with no access to their trusted AI assistant Claude, our volunteer researchers did not expect to be attacked by a four-legged robot.
Yet as the mechanical whirring and rubberized footfalls grew louder, the humans startled. They had been trying, without success, to establish a connection between their computers and a robotic quadruped—a “robodog.” Meanwhile, the competing team on the other side of the room had long since done so and were now controlling their robot with a program largely written by Claude. But in an all-too-human error of arithmetic, Team Claude had instructed their robodog to move forward at a speed of one meter per second for five seconds—failing to realize that less than five meters away was the table with the other team.
The robot did as it was instructed, careening toward the hapless coders. The event’s organizer managed to grab hold of the robot and power it off before any damage was done to robots, tables, or human limbs. The morale of the inadvertently targeted team, however, did not escape unscathed.
At this point, you might be asking…
A common question about the impact of AI is how good it will be at interacting with the physical world. Even as we enter the era of AI agents—which take actions instead of just providing information—these actions are largely digital, such as writing code and manipulating software. We’ve previously explored how AI can bridge the digital-physical divide in a limited way with Project Vend, where we had Claude run a small shop in Anthropic’s office.
In that experiment, AI’s interaction with the real world was mediated by human labor. In this robodog experiment, we took a natural next step and used robots instead of people to tackle a different challenge.
One way of understanding and tracking the capabilities of AI models is to run an “uplift” study. These experiments randomly divide participants into two groups—one with access to AI and one without—and measure the difference in task performance between them (we’ve used this methodology extensively in our work on AI and biological risk). The difference between the groups is the “uplift”—the advantage (if any) provided by AI. Measuring uplift tells us about the present ability of AI to augment human performance. It’s also suggestive of the future domains in which AI will be able to successfully perform tasks on its own.
To run our experiment, we recruited eight Anthropic researchers and engineers, none of whom had extensive prior experience with robots.1 We randomly selected four to be on “Team Claude” and four to be on “Team Claude-less.” Then, we asked each team to operate a quadruped robodog in three increasingly difficult phases. In all phases, the core task they were being evaluated against was simple: get the robodog to fetch a beach ball.
We do not expect robotic fetch to prove so economically valuable that it shows up as a task on a future version of our Anthropic Economic Index. So why are we doing this?
First, it builds on our previous research. One of the evaluations we use to assess the ability of Claude to contribute to AI R&D is a test of its ability to train a machine learning model that could be used to control a quadruped robot. We’ve previously evaluated the resulting algorithm using simulations, which have shown that Claude is not yet at the point where it can handle this task truly autonomously.2 This meant that this task was well suited to a trial that combined AI with human help. We could also be confident our experiment would be useful to repeat in the future: there is still a lot of room for models to improve on robotics.
Another reason is practical. It’s hard to pull our colleagues away from work for more than a day, so we needed a task that was difficult enough to fill that time, but not so difficult that teams would make minimal progress and we would be unable to detect uplift even if it were there. Beach ball retrieval, especially the more difficult variants, met these criteria.
In Phase One, teams had to use the manufacturer-provided controller to make their robodog bring the ball back to a patch of fake grass. This was purely to give the teams a feel for the hardware and what it could do: we didn’t expect any uplift here.3
Phase Two required teams to put down their controllers. They had to connect their own computers to the robodog, access data from its onboard sensors (video and lidar), develop their own software program for moving the robot around, and then use that to retrieve the ball. This is where we expected Claude might begin to provide an advantage.
Phase Three was even harder. The teams needed to develop a program that would allow the robodog to detect and fetch the ball autonomously—that is, without being directed towards the ball by human control. Again, our expectation was that Claude would prove helpful.
Overall, Team Claude accomplished more tasks and completed them faster on average. In fact, for the tasks that both teams completed, Team Claude succeeded in about half the time it took Team Claude-less (see Figure 1). That is: AI provided substantial uplift for this set of robotics tasks.
The task-by-task breakdown of results (split into the three phases) shows where Claude was most advantageous.
The most striking advantage provided by Claude was in connecting to the robot and its onboard sensors. This involved connecting to the dog with a laptop, receiving data, and sending commands. There are a number of different ways to connect to this particular robot, and a lot of information (of varying accuracy) available online. The team with Claude was able to explore these approaches more efficiently.
Team Claude also avoided