Skip to content
Bud's Wiki

Claude Opus 5 driving robot arms zero-shot via Inspect Robots (Jay Chooi / Robocurve)

TL;DR

Section titled “TL;DR”

Jay Chooi (Robocurve; AI safety, Harvard/MATS/Rhodes) posts a short thread of a Claude Opus 5 harness driving physical robot arms zero-shot on two tabletop manipulation tasks: stacking bowls and placing a fork on a plate. The setup logs per-episode wrist camera, context camera, and joint positions, and appears to invoke Opus 5 as the controller without any robotics-specific fine-tuning (“Opus did it zero-shot. We’re surprised, given that LLMs are not predominantly trained for robots”). The evaluation harness — Inspect Robots, MIT-licensed at github.com/robocurve/inspect-robots — is announced as an open-source framework for evaluating VLA and physical-AI models on real hardware, and Chooi flags an upcoming full report benchmarking frontier LLMs and VLAs on real-world tasks at Robocurve.

Key claims

Section titled “Key claims”
  • Claude Opus 5, invoked as a robot controller with wrist cam + context cam + joint positions logged per episode, stacks bowls and places a fork on a plate zero-shot with no robotics-specific fine-tuning [tweet 1, tweet 4].
  • Thin flat objects (fork on plate) are explicitly called out as a difficulty case that the Opus 5 harness still handles [tweet 2].
  • Inspect Robots is released as a free, open-source (MIT) evaluation framework for VLA / physical-AI models, positioned as the harness behind the demos in this thread [tweet 5, github.com/robocurve/inspect-robots].
  • Robocurve is running a broader benchmark of frontier LLMs and VLAs on real-world tasks, with a full report deferred [tweet 6].
  • Aris Zhu (reply): initial impression is that Opus 5 is faster on these tasks than Fable 5 on high effort [reply].

Method

Section titled “Method”

The thread is a product/eval announcement, not a paper — no architecture or training details for the LLM-as-controller stack are given. What can be extracted from the four attached videos and captions:

  • Two tabletop tasks on a dual-arm (or single-arm from wrist-cam POV — unclear) rig: bowl-stacking and fork-on-plate.
  • Two camera streams (a wrist camera and a scene “context” camera) plus joint positions are logged for every episode — the standard observation set used in filed LLM-as-VLA-supervisor work such as Anthropic’s Frontier Red Team evaluation (How Claude Performs on Robotics Tasks).
  • The invocation is described as zero-shot, i.e. Opus 5 is not fine-tuned for robotics; the harness (Inspect Robots) presumably marshals the perception/action interface. Whether Opus 5 emits low-level joint targets, end-effector deltas, or high-level plans over a lower controller is not disclosed in the tweet.
  • Inspect Robots itself is described as “an evaluation framework for VLA / physical-AI models” (from the GitHub OG-card excerpt) and is Apache/MIT-adjacent (MIT per the tweet).

Results

Section titled “Results”
  • Successful zero-shot rollouts on two tasks (bowl-stack, fork-on-plate) shown as ~5–10 s video clips; no quantitative success rate, no per-task N, no comparison against a baseline in the thread itself.
  • Qualitative claim from a reply that Opus 5 is faster than Fable 5 “on high effort” for these tasks — again anecdotal, no numbers.
  • Numbers, model comparisons, and cross-task generalization are explicitly deferred to a promised full report.

Why it’s interesting

Section titled “Why it’s interesting”

Two things make this worth filing despite being a launch tweet with no numbers. First, it is a second frontier-lab-adjacent datapoint on the “frontier LLM as robot controller” surface within a two-week window, complementing How Claude Performs on Robotics Tasks — Anthropic’s own Frontier Red Team evaluation which ran Opus 4→4.7 as controllers over MolmoAct on LIBERO and found every LLM-guided VLA underperformed MolmoAct alone. Chooi is now claiming Opus 5 works zero-shot on real hardware for the specific manipulation subclass (bowl-stacking, thin-object placement) that Anthropic’s report flagged as the ceiling case; the promised Robocurve report will be the first outside test of that generation on real physical tasks. Second, Inspect Robots is a new evaluation harness aimed squarely at the LLM-as-VLA-supervisor interface — sitting in the same design-pattern slot as SETA/Toolathlon-GYM/OpenReward (RL Environment Platforms) but for physical-AI models rather than terminal or MCP-tool agents. The tweet is a pointer, not the artifact; the artifacts to watch are (a) the promised full report and (b) the robocurve/inspect-robots repo, neither of which is filed yet.

See also

Section titled “See also”