01What happened
The story, straight
Anthropic published 'Project Fetch: Phase Two,' a research update on its experiment using Claude to control an off-the-shelf robotic quadruped. In August 2025, Claude Opus 4.1 could not complete the tasks autonomously—it got stuck even connecting to the robot. The new phase shows how rapidly AI capabilities have advanced in under a year, with researchers Michael Ilie, C. Daniel Freeman, and Kevin K. Troy noting that AI models are 'moving fast—even faster than the runaway robodog that almost rammed into one of our human teams back in August.'
anthropic's back with 'Project Fetch: Phase Two' — the experiment where claude tries to make a robot dog do tricks. last august, the model couldn't even figure out how to connect to the thing. now they're saying the gap between 'can't connect' and 'autonomous control' closed in under a year. the researchers — michael ilie, daniel freeman, kevin troy — framed it as a speed-of-progress story. there's also a detail about a runaway robodog almost plowing into a human team, which is objectively funny.
02Spread timeline
Where it actually started
03Source receipts
Every claim, linked
04What's solid, what isn't
What's solid and what isn't
- Anthropic ran Project Fetch Phase One in August 2025 using Claude Opus 4.1 and a commercial robotic quadruped.
- In Phase One, Claude could not complete the tasks autonomously and got stuck connecting to the robot.
- Phase Two, published June 21, 2026, demonstrates significant improvement in Claude's ability to control the robot.
- The research team includes Michael Ilie, C. Daniel Freeman, and Kevin K. Troy.
- The specific capabilities Claude can now perform autonomously in Phase Two (the excerpt is truncated).
- Whether the robodog incident involved physical contact or was a near-miss only.
05Why it matters
The editorial take
The update is notable less for the robodog gimmick and more as a concrete benchmark of how fast frontier AI models are closing capability gaps. The August 2025 baseline—where Claude couldn't even initiate a connection—versus the current phase suggests Anthropic is using Project Fetch as an internal yardstick for autonomous agent progress. The project also highlights the broader trend of AI labs moving from chatbot demos toward embodied, real-world robotics applications.
the robodog is the hook but the real story is the speed. going from 'can't connect to the robot' to 'autonomous control' in under a year is a concrete benchmark, not a press-release handwave. anthropic is basically using a warehouse and a quadruped as a progress meter for claude's real-world agency. and if your AI lab isn't testing against runaway robots, are you even trying.