Project Fetch: Can Claude train a robot dog?

Source: Anthropic Research Date: 2025-11-12 URL: https://www.anthropic.com/research/project-fetch-robot-dog

Summary

8 non-robotics-expert researchers split into Claude vs. no-Claude groups, tasked with programming quadruped robots to fetch beach balls across three difficulty phases. Team Claude completed tasks in ~half the time and accomplished more objectives. Claude’s largest advantage was in connecting to the robot and exploring sensor interfaces. Team Claude wrote 9x more code; Team no-Claude asked teammates 44% more questions.

Implications

Bridges the long-horizon agent and physical world threads. The 9x code volume result is the interesting signal: AI assistance doesn’t just speed up existing approaches, it expands the solution space explored. The 44% more peer questions in the no-AI group suggests AI partially replaces the knowledge-sharing function of collaboration — with uncertain longer-term effects on team knowledge formation. The “uplift precedes autonomy” framing is explicit RSP language — Anthropic is using Project Fetch as evidence for why capability thresholds in the Responsible Scaling Policy need to account for human-AI collaboration, not just autonomous performance. Watch for this framing in Anthropic’s regulatory submissions.