Michael Ilie, C. Daniel Freeman and Kevin K. Troy argue in Anthropic that Claude is taking on more robot tasks

Michael Ilie, C. Daniel Freeman and Kevin K. Troy argue in Anthropic that Claude is taking on more robot tasks

Michael Ilie, C. Daniel Freeman and Kevin K. Troy describe Project Fetch Phase Two, published June 18, 2026. They report that Claude Opus 4.7, working without human assistance, completed tasks that human teams had done with a robotic quadruped at least ten times faster. The test excluded using a physical controller and did not measure how long a researcher would take to use Claude’s programmed controller to retrieve the ball. The authors do not argue that language models have solved robotics: Claude struggled to move the beach ball precisely, which required adjusting as it moved, and its attempts at autonomous retrieval failed. They see a shift from models helping people, to people helping models, to models doing more independently. For people using off-the-shelf robots, they say, this could mean quicker preparation before taking control themselves; for some tasks, a human controlling the robot may still do better. They describe a possible early era of physical agentic AI, while saying more research is needed on customized tools and broader capabilities. Their conclusion is that rapid progress in software makes it unwise to rule out a similar trajectory in hardware. Whether Claude can retrieve the ball with the speed and reliability it showed on other tasks remains unknown.

Publication: fc183043-c387-be4a-ce5a-5755b22882df

attribution — pass: The reading credits the byline authors Michael Ilie, C. Daniel Freeman and Kevin K. Troy and the publisher Anthropic. It keeps the speed results, limits and forecasts as the authors' reports and views rather than established fact. This is a single research post, not a listing.

Michael Ilie, C. Daniel Freeman, and Kevin K. Troy

summary — pass: The reading leads with the actors and the development: Claude Opus 4.7 completed robot tasks faster than earlier human teams. It keeps the main caveats (excluded tasks, the failed precise ball retrieval) and the authors' three-stage pattern. It reports the forward-looking hardware claim as the authors' conclusion. Agentic AI is defined on the page. A robotic quadruped is understandable in plain language. The title and question follow from the source's stated next milestone.

What we will be watching for next, though, is the ability of the models to accomplish this final task with the same speed and reliability they displayed on the other elements of Project Fetch.

date — pass: The reading states June 18, 2026, which matches the UTC calendar day of originalPublished, and the evidence kind is publisher-page. The page shows an update note dated Jun 18, the same day as publication, correcting the Phase One date. It also has a later modifiedAt of Oct 23, 2026. The reading does not rely on the corrected August date or on any later-added passage. It contains no hindsight framing, and it keeps forecasts as the authors' views.

Updated Jun 18: Corrected the date of the first phase of Project Fetch.

support — pass: Each claim's excerpt is an exact substring of the evidence and supports its claim, including the faster completion, the excluded tasks, the failed retrieval and the plausible early era of physical agentic AI. The cited passages are core body text, plainly part of the original piece. The only noted edit, dated Jun 18, concerns the Phase One date, which the reading does not assert. The implied benefit of a quicker move to controlling robots is source-stated. The limitation about three trials matches the source.

which means that people can more quickly transition to controlling and using the robots

media — pass: No media was selected, and no verified media candidates were supplied, so no media is misidentified or irrelevant.