Anthropic’s recent Phase 2 of Project Fetch involved testing how well its AI model, Claude Opus 4.7, can program a robotic dog, also known as a robodog, for autonomous tasks. This phase followed a 2024 experiment where a previous version, Opus 4.1, significantly aided a team of non-experts, enabling them to outperform another team without Claude. In the new phase, Claude Opus 4.7 demonstrated remarkable improvement, completing tasks approximately 20 times faster than the fastest human team from the earlier test. Despite this progress, the model still struggled with the key task of effectively retrieving a beach ball, highlighting the ongoing challenges in integrating AI with physical world applications. This research reflects a broader trend of AI models moving beyond just assisting humans to performing complex tasks independently, a progression observed in areas like cybersecurity and robotics.

Anthropic: Anthropic is an AI research company that develops advanced language models, including the Claude family. In this news, the company conducted Phase 2 of Project Fetch to evaluate how its latest models perform on programming and controlling an off-the-shelf robotic quadruped without robotics expertise.
Claude Opus 4.1: Claude Opus 4.1 is a prior-generation model from Anthropic that was tested in the original Project Fetch experiment. It provided substantial assistance to non-expert human teams in completing robotics programming tasks but required human collaboration and could not handle the full workflow independently.
Claude Opus 4.7: Claude Opus 4.7 is Anthropic’s newest model evaluated in Phase 2 of Project Fetch. Operating with minimal human oversight, it completed the same suite of robodog programming and control tasks far more rapidly than previous human teams assisted by earlier models.

`json
{
“Model Progression”: “AI models are increasingly capable of conducting complex, multi-step tasks independently, transitioning from merely assisting humans.”,
“Research Patterns”: “Significant capability advances seen in domains such as cybersecurity are now appearing at the intersection of AI and robotics.”,
“Physical World Applications”: “Experiments demonstrate language models extending their capabilities to controlling physical devices like robots.”
}
`