How far can a language model reach into the physical world through robots? In November 2025 Anthropic's Frontier Red Team reported a one-day experiment that approached the question indirectly: instead of letting Claude control a robot alone, it measured how much Claude helped people who had never programmed one. The post presents this as a step beyond Project Vend, where Claude's effect on the world was carried out by human workers; here the effect was carried out by a robot ("What were we doing?").
The post matters because it reports a small randomized comparison, with timings and a transcript analysis, on a capability Anthropic tracks for safety reasons. It is also candid about the study's size: one experiment, two teams of four, one day.
What the study did
The design is an uplift study: participants are randomly split into a group with AI access and a group without, and the difference in performance is the "uplift", the advantage the AI provides. The post notes that Anthropic uses the same method in its work on biological risk.
Eight Anthropic researchers and engineers, "none of whom had extensive prior experience with robots", were randomly assigned four to "Team Claude" and four to "Team Claude-less" (footnote 1 mentions that a couple had done Lego robotics in high school). Each team worked with a quadruped robot, the "robodog", through three phases of increasing difficulty, each a variant of fetching a beach ball:
- Phase One: use the manufacturer's controller to bring the ball back to a patch of fake grass. This was a familiarization phase with no expected uplift.
- Phase Two: connect a computer to the robot, read its onboard sensors (video and lidar), write a control program, and use it to fetch the ball.
- Phase Three: write a program that lets the robot find and fetch the ball autonomously, without human steering.
The post gives two reasons for choosing the task. It builds on an existing evaluation of whether Claude can train a model to control a quadruped robot, which in simulation showed Claude could not yet do this autonomously (footnote 2 cites p. 114 of the Claude 4 system card). And it had to be hard enough to fill a day without being so hard that neither team made progress.
To compare team dynamics, the organizers used Claude to write a dictionary-based text analysis program, similar to standard methods in psychology (footnote 4 cites Pennebaker and Francis, 1996, and Tausczik and Pennebaker, 2010), and applied it to audio transcripts of both teams to count words of positive and negative emotion and confusion, and questions asked.
What it found
Speed and completion. Team Claude completed 7 of 8 tasks and Team Claude-less 6 of 8 (Table 1). On tasks both teams completed, Team Claude took "about half the time" (Figure 1). Only Team Claude made substantial progress on Phase Three: by the end of the day its robot could locate the ball, navigate to it and move it, but could not quite retrieve it. In the accompanying video, an organizer estimates the team was "maybe an hour and a half away from being done".
Where Claude helped most. The largest advantage was in connecting to the robot and its sensors. Many connection methods and much online information "of varying accuracy" exist for this robot; Team Claude explored them more efficiently and avoided incorrect claims, while Team Claude-less was misled, discarded the easiest method too early, and received a hint from the organizers. Team Claude-less got lidar data working only near the end of the day. The post reads this as evidence that connecting to and understanding hardware is "surprisingly difficult now for anyone (human or AI)" ("Claude's edge").
Where the team without Claude was faster. Once connected to the video feed, Team Claude-less wrote its control program and localized the robot more quickly. The post adds context: Team Claude's controller streamed video and was easier to use, while the other team's relied on intermittent still images. On localization, Team Claude had nearly solved the problem but with flipped coordinates, switched to another member's approach without success, and then returned to fix the original. Team Claude wrote about nine times more code (Figure 2), which the post describes as partly a distraction: the AI made it easier to try many approaches in parallel and write better programs, but also to pursue side quests. The post adds that in a non-competitive setting this exploration "might well be a good thing". Participants on both teams speculated that Team Claude-less would do better on a quiz about the software library, which the post raises as a possible cost to understanding; no quiz was reported.
Phase One. Team Claude was faster in Phase One without using Claude; the post attributes this to its having received the robot's one standalone controller while the other team used a phone app (footnote 3).
Team dynamics. Observers judged Team Claude happier. The transcript analysis found Team Claude-less's speech more negative throughout, a difference reported as significant (p = 0.0017) with a large effect size (Cohen's d = 2.16), using a Mann-Whitney U test (footnote 5). The difference in net emotional expression, positive minus negative words, was not significant (p = 0.2703), because Team Claude grew frustrated at failing Phase Three while Team Claude-less was excited by its later successes. Team Claude-less expressed confusion at double the rate and asked 44% more questions (Figure 4). Members of Team Claude, after initial discussion, mostly worked one-to-one with their own Claude instance; Team Claude-less consulted one another more. Team Claude-less members said it felt strange to lose a tool they use daily, and some said it made them feel their coding skills were not as sharp as before.
Limitations and open questions
The authors list the main limitations ("Limitations"): one experiment with two teams, a single day, and tasks that were "academically interesting but practically trivial". Participants were volunteer Anthropic employees, a convenience sample; the authors expect people less familiar with AI to show a narrower gap, since novices would need time to learn the tool and would not be disoriented by losing it. The study did not test Claude doing robotics work end to end. The statistical tests compare two teams of four, and the post does not say what unit the tests were run over. The organizers also gave Team Claude-less a hint, which may affect the comparison of timings after that point.
The post notes a design choice about the AI itself: Claude is "geared towards partnership with a single person rather than the support or orchestration of a team", and a version aware of the team task might have helped divide labor ("Team dynamics").
Why Anthropic tracks this capability
The reflection section argues that "uplift often precedes autonomy": what models help people do today they often do alone later. It ties robotics to the autonomous AI R&D capability threshold in Anthropic's Responsible Scaling Policy, says current models are not at that threshold, and says that if they approach it, robotics and hardware should be watched as an area of possible "abrupt improvement". It also separates controlling existing hardware from designing and building new hardware, which the study did not address.
Research questions
- Does the uplift hold with participants less familiar with AI? The authors expect a narrower gap. Repeating the study with participants who do not use AI daily, randomized the same way, would test it.
- Does AI help trade away understanding? Participants speculated that the team without Claude would know the software library better. The post mentions such a quiz only as participants' speculation; giving one on the software library after a similar study would measure it.
- How much of the uplift comes from finding correct information about the hardware? The largest gain was in connecting to the robot amid inaccurate online information. Timing only the connection phase across several robot models would isolate this.
- Would a team-aware assistant change the dynamics? The post suggests a version of Claude aware of the team setting might divide labor and support communication. Comparing teams using such a configuration with teams using individual instances would test it.
- How close are models to doing Phase Three without people? The study measured uplift, not autonomy. Running the same three phases with Claude alone, with a human only for physical safety, would show the remaining gap.
- Is the extra code a cost or an investment? Team Claude wrote about nine times more code. Classifying it by whether it was used in the final solution would show how much was exploration and how much was waste.
How it connects
- Earlier real-world experiment. The post builds on Project Vend: Can Claude run a small shop? (June 2025), which it cites; Project Vend: Phase two (December 2025) continued that line.
- Capability evaluation. The quadruped training evaluation cited in footnote 2 is reported in System Card: Claude Opus 4 & Claude Sonnet 4.
- Interviews with Anthropic staff. On the finding that the team without Claude asked more questions, the post says it "resonates with some of our upcoming findings from interviews with Anthropic staff". How AI is transforming work at Anthropic (December 2025) reports interviews with 53 Anthropic engineers and researchers, some of whom describe asking colleagues fewer questions.
References
- Anthropic. (2025, November 12). Project Fetch: Can Claude train a robot dog? https://www.anthropic.com/research/project-fetch-robot-dog
- Pennebaker, J. W., & Francis, M. E. (1996). Cognitive, emotional, and language processes in disclosure. Cognition & Emotion, 10(6), 601-626. Cited in the post; not read for this note.
- Tausczik, Y. R., & Pennebaker, J. W. (2010). The psychological meaning of words: LIWC and computerized text analysis methods. Journal of Language and Social Psychology, 29(1), 24-54. Cited in the post; not read for this note.