LeRobot demos Claude Code operating SO-ARM101 zero-shot via Anthropic MHS — self-calibration to 4.1mm accuracy

TL;DR

LeRobot (Hugging Face’s robotics arm) participated in Anthropic’s Model Hardware Standard (MHS) research preview and posted a video of Claude Code operating a real SO-ARM101 6-DoF arm with zero training data: no learned policy, no teleoperated demonstrations, no imitation. The agent self-calibrated the workspace by requesting a human to rest the closed gripper on 16 dots the software draws on the camera view (torque dropped, gripper acts as its own probe — no checkerboard or camera intrinsics), then wrote the motion program itself. Reported end-to-end accuracy is 4.1 mm position / 3.0 mm placement, with a best run of 12 bricks placed and all four color groups formed. LeRobot frames this as a research preview with open-source release forthcoming.

Key claims

Claude Code runs a real SO-ARM101 with no policy, no teleoperation, no demonstrations — the motion program is agent-authored end-to-end [tweet body].
Calibration is self-conducted by the arm as its own ruler: torque drops, a human rests the closed gripper on 16 dots the software draws in the camera view, and the robot reads back their positions — no checkerboard, no camera intrinsics required [tweet body].
Reported accuracy: 4.1 mm position, 3.0 mm placement; best run placed 12 bricks with all four color groups formed [tweet body].
The demo is part of Anthropic’s MHS research preview; open-source release is announced as forthcoming but not yet available [tweet body].
Framing: the calibration procedure — not the manipulation — is presented as “the interesting part” [tweet body].

Method

The tweet is the primary artifact; there is no accompanying paper or code drop. The described procedure has two phases. Phase 1 (self-calibration): Claude Code renders 16 target dots as an overlay on the arm’s camera feed; the arm drops motor torque so a human can passively guide the closed gripper onto each dot; the agent reads back joint states at each contact and solves for the workspace-to-image mapping. This removes the standard checkerboard-plus-intrinsics camera-calibration step — the gripper acts as a probe, and the arm’s own proprioception provides ground-truth positions. Phase 2 (task execution): the agent writes a brick-placement program directly in code, without a trained policy or teleoperated demonstrations. Video shows 12 bricks placed across 4 color groups in the best reported run.

Results

Numbers reported in the tweet: 4.1 mm position accuracy, 3.0 mm placement accuracy, 12 bricks placed on the best run, all four color groups formed. No standardized benchmark, no ablations, no failure-mode breakdown. The tweet promises “a full hands-off run, start to finish” as the next milestone, implying the current best run still involved some human touch (the calibration itself is by design human-in-the-loop). The clip is a research preview under Anthropic’s MHS; open-source release is announced but not yet posted.

Why it’s interesting

This is the first concrete deployment demo of Previewing the Model Hardware Standard — Anthropic’s Model Hardware Standard — on the wiki, and it lands squarely on the “explore then compile to deterministic script” operational pattern the MHS post described in the abstract. Under MHS, LeRobot exposes the SO-ARM101 as a driver with read/write primitives plus a natural-language tag layer; Claude Code is the agent that discovers the arm, self-calibrates, and writes the motion. The calibration trick — arm as its own ruler, torque-off gripper as probe — is the kind of empirical device semantics MHS’s tag layer is supposed to encode once discovered, and it side-steps the checkerboard/intrinsics prerequisite most robotics-in-a-lab papers still assume. This is a live counter-example to the VLA Models page’s framing that useful manipulation requires a trained policy: here there is no VLA, no imitation, no RL, and the arm still places 12 bricks — echoing How Claude Performs on Robotics Tasks‘s “Claude as supervisor” thread but pushing one abstraction lower (Claude operates the driver directly, not a VLA that operates the driver). Contrasts with Claude Opus 5 driving robot arms zero-shot via Inspect Robots (Jay Chooi / Robocurve) (Claude Opus 5 zero-shot driving robot arms via Inspect Robots on Robocurve) and Gemini 3.7 Flash saturates Robocurve physical tool-use benchmark at 92% vs 32% for 3.6 Flash (three weeks apart) (Gemini 3.7 Flash 92% on Robocurve tool-use): those are LLM-directly-drives-arm too, but through a benchmark harness. The MHS + LeRobot combination adds a shipping dimension — the driver spec, the vendor buy-in, the promised open-source release — that a benchmark run doesn’t. The open question this demo raises for the wiki: at what accuracy floor does “agent-writes-motion” stop competing with a trained policy? 4.1 mm position is coarse compared to teleop-trained VLA numbers on similar hardware, and the “calibration is the interesting part” framing suggests the authors see the self-description surface as the moat, not the motion quality.