Explore the highlights, challenges, and implications of Project Fetch phase two, where Claude outperformed human teams in robotics tasks.
In a noteworthy update to their previous experiment, Anthropic has released the findings from phase two of Project Fetch, showcasing significant advancements in AI and robotics. This phase explored the capabilities of the latest AI model, Claude Opus 4.7, by evaluating its ability to operate a robotic quadruped, referred to as a robodog. With promising results, the study highlights the progress of general-purpose AI models in performing complex robotics tasks.
Initially launched in August 2025, Project Fetch aimed to assess how effectively Claude could assist Anthropic employees in tackling sophisticated tasks involving a robodog. The experiment revealed that the Claude-enabled team substantially outperformed another team relying purely on traditional resources and manual ingenuity. Claude Opus 4.1 demonstrated its capabilities, but it could not independently manage fundamental tasks, such as connecting to the robot.
As AI technology has advanced rapidly, Anthropic decided to revisit this experiment. The latest development, Claude Opus 4.7, exhibited remarkable improvements in efficiency and effectiveness. Even without human assistance, it executed tasks approximately 20 times faster than the fastest human teams from the previous phase.
The updated experiment focused on validating whether the advancements in Claude's architecture could enable it to outperform earlier models without the need for human intervention.
During the trials, Opus 4.7 was directed to complete the same tasks performed by human teams in phase one, but with impressive speed. The model was capable of accomplishing tasks at least ten times faster than any team from the original phase. For the four tasks both human teams had completed, Opus 4.7 showed an average speed that was over 37 times faster than the team without Claude and slightly above 18 times faster than the Claude-assisted team.
Moreover, the AI's ability to generate effective code on the first attempt was significantly enhanced compared to human efforts in the first phase. It produced nearly ten times less code while achieving at least equal success rates. This improvement reflects Claude's advanced capability to quickly identify optimal paths and approaches in executing robotics tasks. However, Claude's performance was not without its faults. It relied on an outdated object detection algorithm at times, which led to slower execution in certain trials.
Despite Claude's advancements, significant challenges remain in harnessing AI for robotics applications. Even though the AI excelled at many tasks, it struggled with the nuanced and complex task of autonomously fetching and controlling the robodog's movements to retrieve a beach ball. Much like their human counterparts, Claude showed limitations in adapting its approach effectively when faced with obstacles.
Although a researcher with more robotics experience successfully programmed the robodog for autonomous fetching, Claude still requires more nuanced programming capabilities to replicate the task's complexity. This observation reiterates that while AI can significantly enhance human productivity in robotics tasks, it still lags behind in executing tasks requiring high precision.
The implications of Project Fetch extend beyond mere performance metrics. Anthropic's research suggests that as AI models continue to evolve, they may reach a point where they can deploy physical tools more efficiently. The latest findings indicate a transition towards a world where AI can engage with off-the-shelf tools, enhancing physical capabilities and streamlining robotic interactions.
This evolution mirrors the progression seen in software, where AI models successfully transitioned from basic coding tasks to more complex programming roles. Just as AI began mastering coding tools, we could witness similar developments in physical applications, wherein models learn to manipulate hardware in progressively sophisticated ways.
However, understanding AI's full potential in robotics will take additional research. Exploring models’ capabilities to craft specialized control policies will be central in realizing the broader vision of adaptable language models in robotics. While current boundaries exist, the rapid advancements highlight the feasibility of overcoming these hurdles. What seemed impossible a few years ago is now becoming a reality.
As this research progresses, not only will AI play an integral role in supporting human tasks, but its eventual autonomy in handling physical tasks also poses intriguing questions. These advancements could revolutionize industries by allowing robots to perform complex tasks previously thought to require human input.
As AI continues its rapid evolution, the findings from Project Fetch phase two prompt optimism about the future of human-robot collaboration. Anthropic's work provides a glimpse into a world where AI can seamlessly assist or even replace humans in certain contexts, setting the stage for increased automation across various sectors.
The potential for Claude, and similar models, to transcend basic programming and robotics tasks is not just exciting but essential as businesses increasingly seek innovative solutions to bolster productivity and efficiency. The future may hold a landscape where AI's hands-on role in robotics becomes commonplace, thus gradually transforming how we interact with technology.
What is Project Fetch?
Project Fetch is an initiative by Anthropic that evaluates how AI models, specifically Claude, can assist employees in using robotic systems to perform complex tasks.
How did Claude Opus 4.7 perform in the latest phase?
Claude Opus 4.7 demonstrated remarkable efficiency, completing tasks significantly faster than the human teams in the previous project phase, often more than 20 times quicker.
What challenges does Claude face in robotics tasks?
Despite its advancements, Claude still struggles with nuanced controls and high-precision tasks, indicating the necessity for further improvements in programming capabilities.