We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hacker News·Jun 21·5 min read
Project Fetch: Phase Two
TL;DR
In August 2025, a gap existed in using AI to assist non-experts with robotic tasks. Project Fetch utilized the Claude Opus 4.7 model, which significantly improved task performance.
✦ Why It Matters
Engineers can leverage advanced AI models to enhance robotic task performance, while recognizing current limitations.
Key Takeaways
- Explore AI models to assist in complex robotic tasks.
- Recognize the limitations of current AI in precise control.
- Utilize advancements in AI to improve task efficiency.
- Stay informed about the evolving dynamics between AI and robotics.
Full Summary
In August 2025, Anthropic conducted Project Fetch to explore how AI could assist non-experts in performing complex tasks with a robotic quadruped, or robodog. Initially, the Claude Opus 4.1 model was tested, revealing that while it aided human teams, it could not operate the robot independently.
With advancements, the team revisited the project using Claude Opus 4.7, which demonstrated remarkable improvements, completing tasks 20 times faster than the fastest human team from the previous experiment. Despite this progress, the model still struggled with specific tasks like accurately moving a beach ball, indicating that challenges in low-level robotic control persist.
This pattern of AI first assisting humans, then being assisted by them, and finally operating independently is becoming evident in both cybersecurity and robotics. These findings suggest a growing capability of AI in the physical world, but also highlight areas needing further development.