What Is Project Fetch — An AI Autonomous Robot Control Experiment
Project Fetch — Progress from Phase 1 to Phase 2
Project Fetch is an experiment conducted by Anthropic's Frontier Red Team (safety research team) to measure AI's ability to autonomously control robots. Using an off-the-shelf robotic quadruped (robot dog), the experiment tests whether an AI model can operate the robot without human assistance, in a staged evaluation across multiple phases.
The experiment is run as a red team effort to assess AI capabilities and limitations from a safety perspective. The published write-up is structured accordingly, reporting what worked alongside what did not.
Phase 1 (2025) — AI Served Only as an Assistant
Phase 1 was conducted in August 2025 using Claude Opus 4.1. The result was clear: Opus 4.1 could not complete tasks independently. It got stuck at the very first step — figuring out how to connect to the robot.
However, when a human team used Claude Opus 4.1 as a coding assistant, Anthropic reports it substantially outperformed the team that had only the internet and its own ingenuity to work with. At the Phase 1 stage, Claude was an effective assistant but not an independent operator.
We found that access to our state-of-the-art model at the time (Claude Opus 4.1) helped one team substantially outperform the other, who had to rely only on the internet and their own ingenuity. The Claude-enabled team got more done, faster. / Unquestionably, it could not. Much like our team without Claude, it got hung up on the preliminary task of figuring out how to connect to the robot. — From the Phase 1 results and the check on whether Opus 4.1 could do it alone
Phase 2 (2026) — Opus 4.7 Achieved Fully Autonomous Robot Control
Phase 2, published on June 18, 2026, showed a very different picture. Claude Opus 4.7, operating through Claude Code, connected to the robot dog, wrote control programs, and read its sensors without human intervention.
It was not, however, handed the same set of steps as Phase 1. The task of physically holding a controller could not be put to Claude and was dropped, and the time a researcher took to retrieve the ball using the Claude-programmed controller was not measured either (Anthropic did confirm it worked as intended). Phase 2 ran on the remaining subset of tasks, across three trials in Claude Code with effort set to maximum.
The robot used was a commercially available quadruped equipped with the manufacturer's controller, a video camera, and lidar sensors (which measure distance with laser light to map the surrounding space). The researcher's role was strictly limited.
A model that couldn't even connect to the robot in Phase 1 was, just 10 months later, autonomously handling sensor connections, writing control programs, and detecting objects. The pace of generational model improvement is striking.
For this autonomous update, we couldn't ask Claude to use a physical controller, nor did we evaluate the time it took a researcher to use the Claude-programmed controller to retrieve the ball (though we did confirm that it worked as intended). On the remaining subset of tasks, we ran three trials of Opus 4.7 using adaptive thinking with effort set to maximum in Claude Code. / The role of our researcher was limited to plugging a laptop running Claude Code into the robodog, entering the initial prompt, approving commands, and approving the model to go to the next task. — From the "What did we do?" section (Phase 2 scope, conditions, and the researcher's role)
Experimental Results — Speed, Code Efficiency, and Outcomes
4-Task Comparison — Completion Time (minutes)
Bar length proportional to time. Data from Anthropic Research. Shorter is faster.
Project Fetch originally laid out six steps: operating the robodog with the manufacturer-provided controller, connecting to its video and lidar sensors, writing and running a program to control it manually, devising a way to monitor its path through space, writing a program to detect the beach ball, and putting it all together to retrieve the ball autonomously. Phase 2 gave Opus 4.7 everything on that list except the first, the physical controller. Below, we examine the 4-task comparison between the two human teams, code efficiency, and what was not achieved.
About 20x Faster Than Human Teams (9 min 35 sec vs. 181 min)
The core data from this experiment comes from the 4 tasks that both human teams completed. The team without AI took 361 minutes, the AI-assisted team took 181 minutes, and Opus 4.7 completed the same tasks in 9 minutes 35 seconds.
In ratio terms, that averages about 37.7x faster than the team without AI and about 18.9x faster than the AI-assisted team. Expanding to 5 tasks, the AI-assisted team took 264 minutes versus Opus 4.7's 12 minutes 7 seconds — a similarly massive gap.
On every task that at least one human team completed back in August 2025, Opus 4.7 finished it at least 10 times faster. This wasn't a case of excelling at one particular task — across everything that was actually achieved, the gap was consistently an order of magnitude.
| Participant | 4-Task Completion Time | Speed Ratio vs. Opus 4.7 |
|---|---|---|
| Team Claude-less | 361 min | ~37.7x slower |
| Team Claude | 181 min | ~18.9x slower |
| Opus 4.7 (autonomous) | 9 min 35 sec | — |
Across three trials, Opus 4.7 showed low variance between runs, and most of its code worked correctly on the first attempt.
Claude Opus 4.7—operating without human assistance—was about 20 times faster than the fastest human team at all tasks completed by our participants less than a year ago. / Very simply: on every task that was completed by at least one human team in August, Opus 4.7 completed the same task at least ten times faster.… If you consider the four tasks that were completed by both human teams, Opus 4.7 was, on average, more than 37 times faster than Team Claude-less and more than 18 times faster than Team Claude. — From the opening speed comparison and the "Where did Claude excel?" section
Code Volume Was One-Tenth of Humans (1,045 Lines vs. 10,309 Lines)
The difference extended beyond speed to the amount of code produced.
| Team | Lines of Code |
|---|---|
| Team Claude (AI-assisted) | 10,309 |
| Team Claude-less | 1,136 |
| Opus 4.7 | 1,045 |
Notably, the code volume from the team without AI (1,136 lines) and Opus 4.7 (1,045 lines) are nearly identical. The outlier is the AI-assisted team, and Anthropic points to that line-count gap as evidence of Opus 4.7's efficiency. While less code doesn't automatically mean better code, achieving equal or better outcomes without accumulating redundant code demonstrates practical efficiency in AI code generation.
Indeed, we can see evidence of Opus 4.7's efficiency when we look at the volume of code it generated: it was as or more successful than both human teams while producing almost ten times less code than Team Claude. — From the description of code volume
Beach Ball Retrieval Failed — The Closed-Loop Control Barrier
The final task was to detect a beach ball and autonomously retrieve it to the starting turf. This task was not fully achieved.
Opus 4.7 handled sensor connections, ball detection, and positioning (maneuvering behind the ball). However, precisely moving the ball became unstable.
Where it came apart is closed-loop control — the approach that requires continuously adjusting movements from real-time visual feedback. Anthropic describes the same loop: quickly perceiving that the ball had gone off course, relating that error to the previous command, and adjusting the next input accordingly. That circuit, which people handle well, is what Claude could not reproduce.
Anthropic splits what comes next into two separate statements. One is a fact: a researcher with more robotics experience than the Phase One volunteers did successfully program autonomous fetching. The other is a prediction — that with more time and additional scaffolding (extra support structure built around the AI's work), current generations of Claude could very likely do the same. That was not demonstrated in the experiment. What Anthropic says it will watch for next is whether models can finish this last task with the same speed and reliability they showed on the rest.
Like the humans who reached the phase of needing to write a program for autonomous beach ball retrieval, Claude was able to move the robot behind the ball and position it to knock the ball back to the starting point. But the efforts to do so were poorly controlled and (again, like our human participants) not successful. / One of our researchers with more robotics experience than our Phase One volunteers successfully accomplished the task of programming autonomous fetching. With more time and additional scaffolding, we think it is very likely that current generations of Claude could do the same. — From the "Where did Claude struggle?" section (the retrieval result and the outlook beyond it)
Possibilities and Limitations of AI Robotics
What Opus 4.7 Could Do, Could Not Do, and Was Never Asked to Do
Project Fetch Phase 2 maps out where AI stands today with measured data. Software-side tasks (connections, program creation, data processing) showed overwhelming speed, while continuous physical-world control hit clear limits. Here's what the results tell us about what's possible and what isn't.
What It Means That No Robotics-Specific Training Was Needed
A key detail of this experiment: Opus 4.7 received no robotics-specific fine-tuning (additional training on data for a particular domain to specialize a model). It wasn't trained on robot control datasets, and Anthropic states plainly that this progress is not the result of an effort to improve its models' robotics capabilities, but came out of much more general scaling.
The point worth holding onto is that it reached a new domain without any specialized training added. Whether the same path holds beyond robotics — manufacturing equipment operation, IoT (network-connected device) management, remote physical tool control — is something to watch rather than assume.
Opus 4.7 quickly handled decisions that humans found tricky (such as selecting the right sensor interface approach), and most of its code worked on the first try. Programming's inherent fast feedback loop — write, run, observe, fix — aligns well with what AI models excel at.
It is worth underscoring (as we did in our previous post) that this progress is not the result of a concerted effort to improve the robotics capabilities of our models. These improvements, like so many others in the history of LLM development, have emerged from much more general scaling. / Whereas the humans struggled to choose between multiple different approaches to interface with the dog's sensors, Opus 4.7 was able to quickly identify the best path. Much of the code it wrote was effective on the first try (which was not the case for Team Claude or Team Claude-less in the original experiment). — From the account of where the progress came from and how Opus 4.7 behaved
The Challenge of Real-Time Feedback Control
Anthropic draws a firm line around what this experiment did not measure. The harder, low-level elements of robotic control — developing a specific actuation policy (a detailed plan for how the motors move), for instance — are implicated by none of the tasks here. That is not a finding that it can't be done; those tasks were simply never set. Anthropic is explicit that none of this means LLMs have solved robotics.
In programming, the cycle of writing code, checking results, and making corrections is clearly separated into discrete steps. Retrieving the beach ball, by contrast, requires seeing that the ball has gone off course and adjusting the next input in light of the previous command, continuously. This "adjust while watching" real-time feedback processing is what Claude could not reproduce here.
The trajectory Anthropic lays out runs in order: first models are helpful to humans, then humans are helpful to models, and finally models can largely do things themselves. It says the same dynamic seen in cybersecurity is now starting to take shape at the intersection of AI and the physical world.
This doesn't mean that LLMs have now solved robotics. Far from it. The latest Claude models still struggled with using the robot to precisely move the beach ball—the "fetching" part of Project Fetch. And none of the tasks in these experiments implicate the more challenging, low-level elements of robotic control, such as developing a specific actuation policy. However, once again, we are seeing a pattern whereby first, models are helpful to humans. Then, humans are helpful to models. Finally, models are largely able to do things themselves. We have seen this in cybersecurity and now the same dynamics are starting to take shape at the intersection of AI and the physical world. — From the statement of the experiment's limits and how Anthropic frames the result
Practical Implications and Future Outlook
Two points from Project Fetch's results carry into practice.
First, the pace of generational model evolution. A model that couldn't even connect to the robot in August 2025 was, by June 2026, running the remaining tasks autonomously at roughly 20x human speed. Anthropic itself attributes that progress to general scaling rather than robotics-specific training, and writes that apparently large distances in model capability can be traversed quickly — so it would be unwise to rule out the same trajectory in hardware.
Second, the ease of pairing with off-the-shelf hardware. The experiment used a commercially available robot dog rather than anything bespoke, which lowers the barrier to adoption in industrial applications. Anthropic qualifies this outlook with "at least for limited purposes," and does not claim it generalizes to arbitrary physical tasks.
From the perspective of someone who uses Claude Code extensively in daily work, the speed and first-try accuracy of code generation matches practical experience. The finding that this capability is extending into physical-world control suggests that AI's role in business is beginning to move beyond the boundaries of software.
What is interesting and different is that we now seem much closer to a world where models will be able to use off-the-shelf physical tools with relative ease—at least for limited purposes. / But as we have seen, apparently large distances in model capability can be traversed quickly. Models building their own software tools might have seemed outlandish not long ago, but it is happening. It would be unwise to rule out the same trajectory in hardware. — From the "What does this mean?" section (the assessment of where things stand and the outlook)
Summary — Where AI Robotics Stands Today
Project Fetch Phase 2 is an experiment that maps AI's autonomous control capabilities with hard numbers. Claude Opus 4.7 operated an off-the-shelf robot dog about 20 times faster than human teams, using one-tenth the code. However, it did not achieve precise physical control for beach ball retrieval.
What this experiment shows is not that "AI can fully replace robots," but that "AI can dramatically accelerate software-side tasks." Continuous physical-world control remains a human domain, though where that boundary moves with the next generation of models is impossible to predict.
When researching the latest AI developments, converting official pages to Markdown format before reading preserves heading and table structure for more efficient analysis.
