Project Fetch: Phase Two (www.anthropic.com)

Anthropic has unveiled the results of Phase Two of Project Fetch, revealing significant advancements in the capabilities of its Claude Opus 4.7 model when applied to robotics tasks. In a recent experiment, Claude Opus 4.7 was able to complete tasks with a robotic quadruped—referred to as a robodog—approximately 20 times faster than the fastest human team from a previous trial conducted less than a year ago. While early models provided substantial assistance to human participants, the latest iteration demonstrated a remarkable ability to operate independently, generating effective code more efficiently and with less volume than its human counterparts. Despite these advancements, the study highlights that Claude Opus 4.7 still encounters challenges, particularly with tasks requiring precise physical manipulation, such as autonomously retrieving a beach ball. While the model showed promise by quickly navigating tasks, it occasionally defaulted to outdated methods, suggesting room for improvement. Nevertheless, the project marks a pivotal moment for the AI/ML community as it hints at the emerging potential for large language models to transition into ‘physical agentic AI,’ capable of interacting seamlessly with physical tools. This progress raises intriguing questions about the future capabilities of AI in blending cognitive understanding with physical actions, paving the way for further research in making AI systems more adept in real-world environments.