Anthropic showcased a robot controlled by Claude. The AI model outperformed the company’s employees at working with it. However, the chatbot still failed to successfully complete the experiment’s final task. Anthropic presented the results of the second phase of the Project Fetch experiment, in which the Claude Opus 4.7 AI model learned to operate a commercial quadruped robot on its own. During testing, the model completed a range of tasks without human assistance — connecting to sensors, programming, and navigating the robot — outperforming Anthropic employees’ results. Anthropic first launched Project Fetch in August 2025. At the time, company employees with no robotics experience were split into two teams: one got access to Claude, while the other relied only on their own knowledge and online research. The results showed that the team with Claude worked much more efficiently. At the same time, the model of that generation — Claude Opus 4.1 — could not complete the full set of tasks on its own and got stuck at the robot connection stage. In less than a year, the situation changed dramatically. Claude Outperformed Humans In the new test, Anthropic connected a laptop running Claude Code directly to the robot. The researcher’s role was limited to launching the initial prompt and confirming commands. Claude Opus 4.7 was able to: Connect to the robot’s cameras and sensors Set up access to LiDAR data Write control software Build a route-tracking system Implement object recognition Integrate individual components into a single system According to Anthropic, on every task that at least one human team completed during the first phase of the experiment, Claude Opus 4.7 delivered results at least ten times faster. If to consider the four tasks that both human teams completed, the model proved to be more than 37 times faster than the team without Claude and almost 19 times faster than the team that used the AI assistant. Task completion results. Source: Anthropic . The company noted that the model also produced nearly ten times less code than humans while achieving comparable or better results. Where AI Still Falls Short of Humans Despite significant progress, Claude was unable to successfully complete the experiment’s final task — autonomously returning the ball to the starting point using the robot dog. The model correctly identified the ball’s position and positioned the robot for a push, but it struggled with precise real-time motion control. Anthropic explained that the issue lies in what’s known as closed-loop control, where the system must continuously evaluate the outcome of each action and adjust subsequent commands based on changes in the environment. In scenarios like these, humans still hold an edge over large language models for now. At the same time, one of the company’s experienced roboticists was able to program autonomous completion of this task. According to Anthropic, current versions of Claude could also achieve this result with additional tools and time. Anthropic Sees the Dawn of the Era of Physical Agentic AI The company emphasized that improving Claude’s robotics capabilities was not a standalone development goal. The progress was a side effect of overall scaling and model improvements. Anthropic believes the industry is gradually moving through a familiar development cycle: First, models help people Then, people help models Eventually, models start completing tasks on their own A similar process was previously observed in software development and cybersecurity, where AI agents learned to autonomously work with software tools. Now, a comparable transformation is beginning to unfold in the physical world. “We are plausibly entering the early era of physical agentic AI,” Anthropic noted. Anthropic Continues to Expand Claude’s Capabilities The results of Project Fetch became another step in the company’s strategy to develop autonomous AI systems. Earlier, Anthropic published a large-scale study on the impact of artificial intelligence on the labor market and concluded that, for now, the technology mostly boosts worker productivity rather than leading to mass layoffs. Read more in the article: In addition, after surveying 81,000 Claude users, the company recorded mixed attitudes toward AI: nearly half of respondents reported expanding their own capabilities thanks to the technology, although about 20% expressed concerns about losing their jobs. Also recently, Claude Code and Cowork product lead Kat Wu said that the next stage in the development of artificial intelligence systems will be proactivity — models will be able to anticipate a user’s needs even before they formulate a request.

Source: incrypted

Read Original Source →