> ## Content Index
> Fetch the complete content index at: https://claudefox.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Apple paper: Claude Opus 5.5 runs a robot arm by writing its own code
- URL: https://claudefox.com/apple-paper-claude-opus-5-5-runs-a-robot-arm-by-writing-its-own-code/
- Published: 2026-11-10T08:53:00.000Z
- Updated: 2026-11-20T02:54:11.000Z
- Description: Apple researchers' AMP1 harness lets a general-purpose model drive a robot arm by writing Python. With Claude Opus 5.5 it succeeded in 104 of 108 real-world trials and 82.8% of RoboLab-120 simulations (default instructions), at about $1.39 per simulated episode. How it works and the code.
- Author: hupicake
- Tags: Research

*Updated November 20, 2026: Expanded with how the harness works, real-world setup, cost per episode, limits and the released code.*

On Agentic Manipulation Policies is a paper by thirteen Apple researchers, first posted to arXiv on November 6, 2026 and revised on November 18\. It argues that a general-purpose AI model, put directly in charge of a robot arm, can be more reliable than robot policies trained on recorded demonstrations. The authors call this setup an agentic manipulation policy, or AMP. Their reference system, AMP1, uses no demonstrations and no task-specific training. In real-world tests it ran on Claude Opus 5.5.

## How AMP1 works

AMP1 is a harness: the software between the model and the robot. The model works in a persistent Python interpreter and calls an SDK whose reference documentation is added to its system prompt. A robot server answers those calls. It returns camera images, depth and the arm's state, splits the point cloud into tracked objects, and checks and runs motions.

The model does not get ready-made grasp or place skills in the basic harness. Instead it writes code that produces timed waypoints: a target pose for the gripper's tool point, an arrival time, and open or close. The harness interpolates these into a smooth trajectory and refuses motions that are unreachable or too fast, telling the model why; collision checking is left to the model, which can preview a trajectory before running it. The model also chooses what to look at: an image enters its context only when its own code displays it.

With nothing but these basics, the paper reports the agent developing its own strategies rather than using predefined skills. It searched for hidden objects, used one object as a tool to move another, and pushed items into graspable positions. It also calibrated its own cameras. In one demonstration it won a game of chess as black and then put the pieces away in drawers.

## The results

- RoboLab-120 simulation (120 tasks, 10 episodes per task in each of three instruction phrasings, 3,600 in all): with the default instructions, 85.6% success with Gemini 4 Argon and 82.8% with Claude Opus 5.5\. The best vision-language-action and world-action entries on the leaderboard scored 49.2% and 42.9%. Across the 13 models tested, results ranged from 12.8% (Qwen3.6-35B-A3B) to 85.6%; the best open-weight model, Qwen3.8-Flash-Next, reached 64.9%.
- CaP-Gym suites, zero-shot: AMP1 scored 95.2% on LIBERO-PRO, 100% on robosuite, and 96.0% on two BEHAVIOR-1K mobile manipulation tasks. The paper excerpts we have don't tie these to Opus 5.5, and the repository's reproduction command for these suites uses Gemini 4 Argon.
- Real world, with Opus 5.5 at high reasoning effort: 104 of 108 trials across 12 tasks, or 96.3%. The two learned baselines, FLUX 3 Action and π0.5, scored 45.4% and 27.8%; the paper describes this as cutting the failure rate 20 and 15 times.
- Learning from experience: adding a memory store, a folder of notes and code the agent reads at the start of each episode, cut the failure rate on RoboDojo bimanual tasks by 4.6 times. In the few-shot setting, success rose from 26.4% zero-shot to 84.1% after practice.

The real-world rig was a 7-joint Franka Research 3 arm with a Robotiq 2F-85 gripper, one wrist camera and two shoulder cameras. Each method got nine trials per task, three on each of three instances of the platform that differed in background and hardware details such as calibration, cabling and wear, with a five-minute limit on execution time (reasoning time excluded). Most errors came from collisions with the table, the bin or the arm itself, which trip the Franka controller's collision reflex and stop the arm. The harness lets the agent query the arm and request recovery, then replan. With that, only four trials failed: twice cubes fell outside a plate while pouring, twice a bowl rolled off the table. Without it, AMP1 succeeded in 94 of 108 trials (87.0%).

## What the model choice changes

The paper's controlled tests find that the model matters more than anything else. On RoboLab-120, AMP1 scored 82.8% with Opus 5.5, 51.6% with Haiku 5.5 and 35.2% with Opus 4.5\. More reasoning helped every model. An assistive harness with grasping and placing routines made the four strongest models worse by up to 7.5 points and the six weakest better by up to 17.6\. Stronger models also coped with missing inputs. As the perception tools, depth and calibration were removed in turn, Opus 5.5 went from 82.8% to 81.9%, 81.1% and 75.8%, while model calls per episode rose from 34.8 to 53.5\. Without calibration it first estimated the camera parameters itself from images taken at known gripper poses. With RGB images only, Haiku 5.5 fell to 10.7%. Replacing images with text descriptions cut Opus 5.5 to 36.8%, from 75.8% with plain RGB images (the introduction compares it with the full-sensor 82.8%).

## Cost and speed

At list prices, an Opus 5.5 episode on RoboLab-120 cost $1.39 on average. The authors compare that with $36.40 reported for another agent, Astranomy, running GPT-6 Astra. They put the gap down to AMP1 commanding whole trajectories per turn instead of tiny steps, and loading images only on request. For Opus 5.5, higher reasoning effort added 14.6 points of success for about $0.90 more per episode. The authors are clear about the limits: AMPs are not yet practical, because the robot spends much of each episode waiting for the model to think, and the model cannot reconsider a trajectory while it runs, so the tests largely use quasi-static scenes.

## Try the code

Apple released AMP1 and the evaluation code on GitHub as apple-aiml-research/ml-amp. You need Linux on x86\_64 with NVIDIA GPUs, the uv package manager, and an API key for the model you test. The README's setup, which also clones the third-party RoboLab benchmark, starts like this:

```
git clone https://github.com/apple-aiml-research/ml-amp AMP && cd AMP
git clone https://github.com/NVlabs/RoboLab.git external/robolab
git -C external/robolab checkout 9db0aaf09d9fe5d4f37b168320788258c7012463   # v0.3.1
git -C external/robolab lfs pull   # needs git-lfs
uv sync --frozen
export GEMINI_API_KEY=...   # or ANTHROPIC_API_KEY
uv run amp doctor
```

The README links a quickstart for the full install and notes that the paper's simulations ran on NVIDIA H100 80GB GPUs under Debian 13, so other hardware may give slightly different numbers. Our advice: each episode is billed in API tokens, so try a handful of tasks before a full benchmark run, and never leave a real arm running unattended.

## Our take

The paper shows that the skills Claude users rely on every day, writing code, checking its result and trying again, can drive a physical arm in these tests. The striking result is less the headline score than the ablations. A bare harness worked best for the strongest models, and seeing the scene mattered far more than having it described. The wait for the model to reason is a real barrier, and it is why this stays a research result for now. For anyone building agent tools, the design is worth copying: small, well-documented primitives, a validator that explains refusals, and images loaded only on request.

Source: [On Agentic Manipulation Policies (arXiv)](https://arxiv.org/abs/2611.09892?ref=claudefox.com), [On Agentic Manipulation Policies, full text v2 (arXiv HTML)](https://arxiv.org/html/2611.09892?ref=claudefox.com), [AMP1 project page (Apple)](https://apple-aiml-research.github.io/ml-amp/?ref=claudefox.com), [apple-aiml-research/ml-amp (GitHub)](https://github.com/apple-aiml-research/ml-amp?ref=claudefox.com)

This story first appeared in the [claudefox daily briefing for November 10, 2026](https://claudefox.com/daily-2026-11-10/).