Project Fetch: Phase Two

anthropic.com

June 21, 2026

7 min read

🔥🔥🔥🔥🔥

47/100

Summary

Project Fetch involved an experiment in August 2025 where Anthropic employees utilized Claude Opus 4.1 to perform complex tasks with a robotic quadruped. The initiative aimed to assess the effectiveness of AI assistance for non-experts in robotics.

Key Takeaways

Claude Opus 4.7 completed tasks related to operating a robotic quadruped 20 times faster than the fastest human team from the previous Project Fetch experiment.
The latest Claude models still face challenges with precise robotic control, particularly in tasks like fetching a beach ball.
The experiment demonstrated a progression where AI models first assist humans, then humans assist models, and eventually models operate independently.
Opus 4.7 was able to effectively write code for robotic tasks on the first attempt, unlike human teams which struggled with multiple approaches.

Read original article

View story →

Coverage & discussions

Hacker News · anthropic.com

Project Fetch: Phase Two ↗

Community discussions

Hacker News discussion ↗

Historical count: 5 · 5 saved in sample · Historical sample

Analysis ready

Community Sentiment

Mixed

Positives

LLMs demonstrate an emergent capability to learn and complete tasks in physical space, suggesting potential for innovative applications in robotics without extensive fine-tuning.

Concerns

The article seems to prioritize speed over robustness, raising concerns about the long-term effectiveness of AI in complex robotics tasks that require expert knowledge.
Preliminary trials with Claude Mythos Preview indicate that the experimental setup may not provide a fair comparison with other models, potentially undermining the reliability of the results.
There are doubts about the maintainability of the code produced by the AI, suggesting that the focus on rapid development may compromise quality.

When AI Builds Itself: Our progress toward recursive self-improvement

Jun 4, 2026