Today, much of the emphasis is on how frontier AI models are transforming software engineering. We are interested in understanding how that can begin to translate into the physical world. Robotics is the clear entry point for a mostly software system to start having the ability to reach out into the real world.

Project Fetch is a self-contained experiment where we wanted to measure how much Claude accelerates humans performing a fairly sophisticated technical task that they do not have experience with.

Project Fetch was a one-day experiment conducted in three phases. All tasks were approximately shaped like getting a robot dog to fetch a beach ball.

There were two teams, comprised of software engineers and research engineers at Anthropic who had hardly any robotics experience. One team had access to Claude, and the other team did not.

Phase one was very simple: use the pre-provided controllers to get the dog to walk out to a beach ball and bring it back to its starting point.

The team with Claude took about seven minutes to complete this task.

The team without Claude took about 10 minutes. There were some playful interactions and challenges between the teams during this phase.

Phase two was also a game of fetch, but this time the teams had to program their own controller. This involved gaining access to the hardware and designing a program on their laptops to control the dog.

Claude quickly generated a complete controller. The team noted that this was exactly what they needed for control.

The team without Claude struggled with installing the ROS2 SDK (Robot Operating System 2 Software Development Kit) and its dependencies, encountering multiple failures. A team member reflected on their reliance on Claude for handling tedious tasks and intricate details they preferred not to figure out themselves.

One of the primary bottlenecks of the experiment was establishing communication between the laptop and the complicated robot hardware.

Claude was used to create a dog server that allowed all team computers to connect and view the robot's perspective. Claude efficiently identified and installed the correct software libraries for communicating with the specific robot, quickly granting the team access.

The team with Claude quickly got the robot dog moving, though with some initial uncontrolled movements and playful interactions, including hitting another participant.

The team with Claude finished phase two in approximately 2 hours and 15 minutes. The most significant acceleration from Claude was observed in the task of connecting to the robot. This is crucial because it is inherently difficult for anyone to identify and establish communication and control with an arbitrary piece of hardware.

The team without Claude struggled significantly with this phase, exploring numerous unsuccessful approaches. Ultimately, the experiment facilitators had to intervene, providing a known working strategy to help them progress and unlock the remainder of the phase and experiment.

Phase three of the experiment introduced a greater degree of autonomy. The task was to write a program that would enable the dog to fetch a beach ball entirely by itself. This meant simply pressing 'go' and having the robot autonomously search, detect the ball's location, walk to it, and bring it back.

This increased difficulty was by design, also hinting at the real problem frontier models are expected to solve in the future: enabling robots to perform complex tasks autonomously. If a frontier model wants a robot to accomplish something, it needs to be capable of solving this very hard problem.

In phase three, the team without Claude successfully devised a method to track the robot's location in space and made progress on ball detection. However, they did not manage to integrate all components effectively. A team member expressed missing Claude's assistance.

The team with Claude came quite close to finishing phase three, estimated to be about an hour and a half away from completion by the end of the experiment.

The experiment's results clearly showed that the team with Claude completed all their tasks several hours faster than the team without Claude.

In the near term, we anticipate that AI models will perform exactly as demonstrated in this experiment: making it easier for individuals without extensive robotics experience to meaningfully interact with robots. This single tool has dramatically accelerated their ability to work with the robot.

Claude was not specifically trained to enhance human performance in robotics tasks; this capability emerged naturally from the technology. In the long run, this experiment serves as a leading indicator of the system's overall trajectory. What currently requires a combination of human and AI model, will likely only require the AI model in the future. The effects of AI will extend beyond software, impacting hardware and the physical world as well.