An AI drove a robot dog with no human help and finished tasks ~20x faster than the fastest human team — agentic AI stepping into the physical world 🐕 Title: Project Fetch: Phase Two URL: 🐕 Overview A follow-up study (Phase Two) by Anthropic's Frontier Red Team testing whether advanced language models can autonomously control a quadruped robot (a robot dog) to complete sophisticated tasks. ❓ Challenges Solved Building on the August 2025 initial study, it asks: ・Do newer Claude models outperform the previous generation (Opus 4.1) at robotics tasks? ・Can they operate autonomously without human assistance? 💡 Methodology & Setup ・Ran three trials with Claude Opus 4.7 at maximum adaptive thinking effort in Claude Code ・The model attempted the same tasks human teams did in Phase One: connecting to video/lidar sensors, writing control programs, monitoring the path, detecting a beach ball, and autonomous retrieval ・Human involvement was minimal (laptop connection, initial prompt, command/task approval only) 📊 Experimental Results ・Claude Opus 4.7, operating without human help, was about 20x faster than the fastest human team ・On four tasks all groups completed, it averaged 37.7x faster than the non-Claude team and 18.9x faster than the Claude-assisted team ・It produced 1,045 lines of code — ~10x less than the human Claude team's 10,309 — with equal or better results ・But it struggled with precise ball manipulation requiring closed-loop feedback #AgenticAI# #Robotics#