Share
Apple has a new robotics paper. Its argument is that you may not need a robot-specific brain to make a robot useful.
The system is called AMP1, short for Agentic Manipulation Policies. It lets general-purpose AI models control robot arms by writing and running Python code. According to the research, it outperformed specialist robot policies on manipulation tasks.
The paper, titled “On Agentic Manipulation Policies” (arXiv:2611.09892), was published in November 2026. The code and simulation data are publicly available.
How AMP1 actually works
AMP1 runs zero-shot. The model gets no fine-tuning on the specific robot or task, and it doesn’t need a library of demonstrations.
The research describes the framework running across several robotic configurations: single-arm, bimanual (two arms), and mobile setups. Supporting different hardware without retraining is exactly the flexibility specialist systems tend to lack.
AMP1 can also adapt its strategy in real time, including using tools and running searches.
The numbers
In simulation, the team tested AMP1 on the RoboLab-120 benchmark. Paired with the Gemini 4 Argon model, AMP1 hit an 85.6% success rate. With Claude Opus 5.5 driving, it reached 82.8%.
For comparison, NVIDIA’s Astranomy model scored 74.8% on the same benchmark.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
We respect your privacy. Unsubscribe anytime.
AMP1 also posted top results on other benchmarks, according to the research. It scored 95.2% on LIBERO-PRO and 100% on robosuite.
The researchers ran 108 real-world trials across three Franka robot arm setups. 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. Much of that improvement is attributed to better error recovery.
Who’s behind it
The study’s authors include Apple researchers Chen Chen and Omid Bagherieh. The work arrives on the heels of Apple’s HALO system for humanoid navigation, which was also released in November 2026.
Background: specialists vs. generalists
The specialist categories AMP1 was measured against include vision-language-action models (VLAs) and world-action models (WAMs). VLAs take in what a robot sees plus a text instruction and output physical actions. WAMs add a learned model of how the world behaves. Both are built specifically for embodied tasks and usually depend on lots of robot-specific training data.
AMP1 takes a different approach. Rather than building a robot brain from scratch, it borrows the reasoning of large general-purpose models and gives them a way to act through code.
What this means
The real-world tests covered 108 trials on Franka arms. AMP1 also 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 public release of code and simulation data means other labs can test these claims directly.
