Apple's AMP1 lets general AI models run robot arms, beats specialists, paper says
In brief
- Apple's AMP1 lets general-purpose AI models control robot arms by writing and running Python code.
- AMP1 runs zero-shot, with no task fine-tuning and no demonstration library.
- AMP1 scored 85.6% on RoboLab-120 when paired with Gemini 4 Argon, according to the paper.
- AMP1 completed 104 of 108 real-world Franka arm trials; baselines finished 49 and 30.
- Results come from a research paper and depend partly on third-party models.
What the paper claims
The paper, titled "On Agentic Manipulation Policies" (arXiv:2611.09892), was published in November 2026, and its code and simulation data were made public. LeoDex News didn't review the paper directly. Every figure below is the researchers' own result as reported by Crypto Briefing (none of it has been independently verified).
AMP1 runs zero-shot. There's no fine-tuning on the specific robot or task, and it doesn't need a library of demonstrations. The research describes it running on single-arm, bimanual (two-arm) and mobile setups, and it can adapt its strategy in real time, including using tools and running searches.
That's the contrast that matters here. The specialists it was measured against, vision-language-action models (VLAs) and world-action models (WAMs), are built for embodied tasks and usually depend on lots of robot-specific training data.
The numbers
In simulation, the team tested AMP1 on the RoboLab-120 benchmark. Paired with Gemini 4 Argon, it hit an 85.6% success rate, according to the paper; with Claude Opus 5.5 driving, it reached 82.8%. The paper's own comparison puts NVIDIA's Astranomy model at 74.8% on the same benchmark. AMP1 also scored 95.2% on LIBERO-PRO and 100% on robosuite, per the research.
The real-world gap was wider.
The researchers ran 108 trials across three Franka robot arm setups, and AMP1 completed 104 of them (a 96.3% success rate). The baseline systems finished 49 and 30 trials, respectively. The paper also reports that AMP1 cut operational failures by 15 to 20 times compared with leading baseline models, and it credits much of that to better error recovery.
The caveats
These are research results from a paper, not a product launch. AMP1 also leans on third-party models such as Gemini 4 Argon and Claude Opus 5.5, so its performance depends partly on systems Apple doesn't control.
The authors include Apple researchers Chen Chen and Omid Bagherieh. The work follows Apple's HALO system for humanoid navigation, which was also released in November 2026.
Frequently asked questions
How does Apple's AMP1 control a robot arm?
AMP1 lets general-purpose AI models control robot arms by writing and running Python code. According to the research, it runs zero-shot, with no fine-tuning on the specific robot or task and no library of demonstrations. It can also adapt its strategy in real time, including using tools and running searches.
How did AMP1 perform against specialist robot models?
According to the paper, as reported by Crypto Briefing, AMP1 reached 85.6% on RoboLab-120 with Gemini 4 Argon and 82.8% with Claude Opus 5.5. The paper's comparison puts NVIDIA's Astranomy at 74.8%. In 108 real-world trials, AMP1 completed 104, while baseline systems finished 49 and 30.
Does AMP1 depend on AI models Apple doesn't own?
Yes. AMP1 relies on third-party models such as Gemini 4 Argon and Claude Opus 5.5, so its performance depends partly on systems Apple does not control. The results come from a research paper, not a product launch.
