The problem
How it works
Anthropic's Frontier Red Team ran a one-day experiment with two randomly assigned teams of four. Both had to get a robot dog to fetch a beach ball in three stages: first with the maker's controller, then by connecting laptops to the robot's sensors and writing their own control software, and finally by building a program that finds and fetches the ball on its own. One team could use Claude; the other could not.
Results and limitations
As vendor research claims, Anthropic reports that the team with Claude completed seven of eight tasks against six, finished shared tasks in about half the time and wrote roughly nine times more code. It was strongest at connecting to the hardware and detecting the ball; the team without Claude was faster at some manual control steps. Only the Claude team got close to autonomous fetching, though its robot was not precise enough to finish.
Before you try it
This is one small experiment with two teams of volunteer Anthropic staff who use Claude daily, run over a single day on tasks the authors call practically trivial, so it is not a general productivity estimate. The post names Claude but not the product or model version. If you try something similar, start with the robot's own safety limits, test new control code at low speed in a clear space with a stop control in reach, and keep people away from moving hardware. Results depend on your setup and were not independently reproduced by Top 200 Repos.
Tools and repositories
No ranked repository is involved. This use case still stands on its own source.
- Tools
- Claude · A programmable robot with an official SDK · A laptop connected to the robot
- Prerequisites
- A robot with a documented programming interface and a safe, clear test area.
Try it yourself: Plan safe first steps with a programmable robot
Fill the fields to adapt the prompt. Empty fields stay as visible [placeholders].
I have a [Robot and model] and want it to [What it should do]. I have no robotics background. Using the maker's official SDK and documentation only, help me plan the work in small stages: connect and read sensor data, then send slow test movements, then add [Next capability]. For each stage give the code, how to test it with the robot at low speed in a clear area, and how to stop it immediately. Do not suggest disabling built-in safety limits or flashing unofficial firmware. Point out anything you are unsure about in the SDK so I can check it in the documentation first.Keep a physical or software stop within reach whenever new code runs.
Sources
- Project Fetch: Can Claude train a robot dog? (opens in a new tab)
Vendor post · anthropic.com · Published 12 Nov 2025 · Source checked 12 Nov 2026, 03:31 UTC
What this source supports
- Eight researchers without robotics experience were split into two teams of four, one with Claude and one without
- Three phases: manual control, connecting to the robot and its sensors with custom software, then autonomous fetching
- Team Claude completed 7 of 8 tasks against 6 of 8 and took about half the time on shared tasks
- Team Claude wrote about nine times more code; only it made substantial progress toward autonomous retrieval
- Limitations: one experiment, small convenience sample of heavy Claude users, a single day of simple tasks