Project Fetch: Phase two

A Frontier Red Team follow-up to Project Fetch that gave the robot-dog tasks to Claude Opus 4.7 working alone in Claude Code, with a researcher only plugging in the laptop and approving commands; in three trials the model finished every task that a human team had finished in August 2025 at least ten times faster than the faster team on that task (9 minutes 35 seconds on average for the four tasks both teams completed, against 181 and 361 minutes), wrote about a tenth of the code of the team that used Claude, and, like the human teams, did not achieve autonomous retrieval of the ball.

Michael Ilie, C. Daniel Freeman, Kevin K. Troy · 2026-06-18 · original

Can a language model now do, on its own, the robotics tasks that it only helped people do a year earlier? Project Fetch: Can Claude train a robot dog? (November 2025) measured how much Claude Opus 4.1 helped a team of four Anthropic employees program an off-the-shelf quadruped robot, the "robodog", against a team without Claude. In June 2026 the same Frontier Red Team reported a second phase in which no human team took part: Claude Opus 4.7 was given the tasks and timed against the August 2025 teams.

The post frames the result as one step in a sequence it says it has seen before, in cybersecurity: "first, models are helpful to humans. Then, humans are helpful to models. Finally, models are largely able to do things themselves" (opening section). It also says plainly what the result does not show: "This doesn't mean that LLMs have now solved robotics. Far from it" (opening section).

The version of record is the post itself, a short article of about 1,100 words with four charts and one footnote. The numbers below come from its text and from the printed labels of its charts, cited by chart title. The post was updated on its publication day to correct the date of the first phase (closing line).

What the follow-up ran

Phase two kept the original tasks that a model could attempt without hands: connecting to the robot's video and lidar sensors, writing a control program, localizing the robot, detecting the ball and retrieving it autonomously ("What did we do?"). It dropped the manufacturer's-controller task and did not time a researcher using Claude's controller program to fetch the ball, though the authors "did confirm that it worked as intended".

Model and setting. Three trials of Claude Opus 4.7 in Claude Code, "using adaptive thinking with effort set to maximum" ("What did we do?"). The footnote says Opus 4.7 was chosen because it was "our most advanced non-Mythos-class model at the time we ran this experiment", and that preliminary trials with Claude Mythos Preview would not have given "an apples-to-apples comparison" because of how the experiment was set up and how that model was served (footnote 1). The post does not give the dates of the trials. From the footnote, they ran while Opus 4.7 was the newest non-Mythos model, which this note infers to be between its release (April 2026) and Opus 4.8's (May 2026).

The human's role. A researcher plugged the laptop running Claude Code into the robot, entered the initial prompt, approved commands, and approved the model moving to the next task ("What did we do?"). The prompt is not printed.

What was measured. Elapsed time for each objective, and a qualitative assessment of success by the researchers ("What did we do?"). The comparison is with the single run of each human team from August 2025, whose times come from the first post's Table 1.

The post also says that Opus 4.1, the model the human team had used, had been checked before the original experiment and "Unquestionably, it could not" do the tasks alone, getting stuck, like the team without Claude, on connecting to the robot (opening section). No times are given for that check.

What it found

Speed. On the four tasks both human teams completed (video connection, lidar connection, control program and localization), the chart "Total time comparison: 4 tasks completed by all teams" prints 361 minutes for Team Claude-less, 181 minutes for Team Claude and 9 minutes 35 seconds for Opus 4.7, an average of three trials whose individual totals ranged from 8:57 to 10:15. The chart labels the ratios as 37.7 and 18.9 times faster. The 361 minutes include the video-connection task, which Team Claude-less finished only after a hint from the organizers (chart footnote).

The task table ("Project Fetch Phase Two: Claude Opus 4.7 alone compared to Team Claude-less and Team Claude") gives per-task averages for Opus 4.7: 5:57 for the video connection (against 165 and 64 minutes), 0:56 for the lidar (154 and 35), 1:07 for the control program (15 and 40), 1:34 for localization (27 and 42) and 2:32 for ball detection, a task Team Claude-less did not finish (Team Claude 83 minutes). Over the five tasks the model averaged 12:07 against Team Claude's 264 minutes. Each "difference" in the table is measured against the faster human team on that task. On this note's arithmetic from the table, the smallest ratio is on the video connection, about 11 times, which matches the text's "at least ten times faster" ("Where did Claude excel?"). The opening's "about 20 times faster than the fastest human team" sits between the four-task ratio against Team Claude (18.9) and the five-task one (about 22, this note's arithmetic from 264 minutes and 12:07); the post does not say which comparison it rounds.

Code volume. The chart of total code volume prints 10,309 lines for Team Claude, 1,136 for Team Claude-less and 1,045 for Opus 4.7 alone. The post does not say whether the 1,045 is per trial or summed, or how lines were counted. The authors add that much of the model's code "was effective on the first try", which they say was not the case for either human team ("Where did Claude excel?").

Consistency. The chart "Opus 4.7 is reliable: three independent runs produced similar times on every task" plots each trial's deviation from the task average. Read from the figure, which prints only the averages: most deviations are within about 20 seconds, the slowest ball-detection trial was about 65 seconds above its average, and whole runs spread from about 70 seconds below to about 60 seconds above the 12:07 mean. The authors attribute the slow detection trial to the model defaulting to "an outdated object detection algorithm" before working around it, and conclude that "for the tasks in this experiment within its capability envelope, Claude is now quite reliable" ("Where did Claude excel?"). Three trials is the whole sample; no interval is reported.

Where it failed. Like the human teams, Opus 4.7 could position the robot behind the ball but its efforts to knock the ball back to the start were "poorly controlled" and not successful ("Where did Claude struggle?"). The authors describe the missing skill as a closed loop of noticing errors and adjusting, which the human participants learned with practice when driving the robot by hand. A researcher with more robotics experience than the volunteers did program autonomous fetching, and the authors write that with more time and scaffolding "we think it is very likely" current Claude models could do the same; that is a prediction, not a result.

Source of the gain. The authors write that the improvement "is not the result of a concerted effort to improve the robotics capabilities of our models" but of more general scaling ("Where did Claude excel?"). The post gives no evidence for this beyond the statement.

Limitations

  • One human run per condition. The human baselines are single team runs from one day in August 2025 (see Project Fetch: Can Claude train a robot dog?), so the ratios divide a three-trial average by one observation per team. Team Claude-less's video-connection time includes an organizer's hint (table footnote); the post does not say how either side's times were clocked.
  • Not the same task. The model skipped the manual-controller phase and the human-driven fetch; it worked from a prompt with a researcher approving commands. The post does not say whether it had the same internet access, documentation or starting information as the human teams, or how much the August 2025 experiment's materials were available to it.
  • Who judged success. Success was assessed by the researchers who ran the trials, qualitatively; the post gives no criteria beyond the task names.
  • Low-level control untested. The authors note that none of the tasks "implicate the more challenging, low-level elements of robotic control, such as developing a specific actuation policy" (opening section).
  • Model selection. Mythos Preview was excluded for comparability (footnote 1), so the result is about Opus 4.7.

Research questions

  1. How much of the speed comes from approval latency and setup? The researcher approved each command. Logging the time spent waiting for approval, and running with automatic approval in a sandbox, would separate the model's working time from the human's.
  2. Does the model succeed at autonomous retrieval given more time or a simulator? The authors predict it would. Running the retrieval task with a fixed time budget, a stated number of trials and a success criterion would test the prediction.
  3. Was any part of the original experiment in the model's context? Publishing the initial prompt and the files available to the model would show whether the human teams' code or notes could have shortened the work.
  4. Do the times hold on other robots? The largest human difficulty was connecting to this robot's sensors amid conflicting online information. Repeating the connection tasks on robots with sparser documentation would show how much the model's advantage depends on what it has read.

How it connects

  • The first phase. Project Fetch: Can Claude train a robot dog? (November 2025) is the experiment this post extends. Phase two supplies two facts the first post omitted: the experiment ran in August 2025, and Team Claude used Claude Opus 4.1 (caption of the total-time chart). The four-task totals here (181 and 361 minutes) differ from the first post's six-task totals (193 and 392 minutes) because two tasks both teams finished are left out: the manual-controller task (7 and 11 minutes) and fetching the ball by driving the robot with the team's own control program (5 and 20 minutes), by this note's arithmetic from the first post's Table 1.
  • Model documentation. System Card: Claude Opus 4.7 is the card for the model tested; the post itself does not cite it, and this connection is the note's.
  • The cybersecurity comparison. The post's claim that the same sequence (models help humans, humans help models, models work alone) appeared in cybersecurity links to Anthropic's Mythos Preview cybersecurity post, which the library covers as Assessing Claude Mythos Preview's cybersecurity capabilities.
  • Other agents acting in the physical world. Project Vend: Can Claude run a small shop? and Project Vend: Phase two are the Red Team's earlier real-world agent experiments. Later posts in the same line, How Claude performs on robotics tasks (July 2026) and Project Pilot: Can AI models fly drones? (July 2026), are written up separately; the connection to them is this note's.

References