Apple researchers have published a new robotics paper with a provocative argument: a robot may not need a purpose-built brain to be useful. The system, called AMP1—short for Agentic Manipulation Policies—enables general-purpose AI models to control robot arms by writing and executing Python code. According to the research, it outperformed specialist robot policies on manipulation tasks in both simulation and real-world testing.

The paper, titled “On Agentic Manipulation Policies” (arXiv:2611.09892), was published in November 2026. The code and simulation data have been released publicly, allowing other labs to verify the claims directly.

How AMP1 Works

AMP1 runs zero-shot. The underlying model receives no fine-tuning on the specific robot or task, and the framework does not require a library of demonstrations. The research describes AMP1 operating across several robotic configurations, including single-arm, bimanual (two-arm), and mobile setups. Supporting different hardware without retraining is precisely the flexibility that specialist systems tend to lack.

The framework can also adapt its strategy in real time, including using tools and running searches as circumstances demand.

The Results

In simulation, the team evaluated AMP1 on the RoboLab-120 benchmark. Paired with the Gemini 4 Argon model, AMP1 achieved an 85.6% success rate. Driven by Claude Opus 5.5, it reached 82.8%. For comparison, NVIDIA's Astranomy model scored 74.8% on the same benchmark. Because AMP1 ran zero-shot, these comparisons pit general-purpose reasoning against specialist policies built specifically for embodied tasks.

AMP1 also posted top results on other benchmarks, according to the research, scoring 95.2% on LIBERO-PRO and 100% on robosuite.

Real-world performance was assessed across 108 trials on three Franka robot arm setups. AMP1 completed 104 of them, a 96.3% success rate, while the baseline systems finished 49 and 30 trials, respectively. The paper additionally reports that AMP1 cut operational failures by 15 to 20 times compared with leading baseline models—an improvement largely attributed to better error recovery.

The Team and the Broader Context

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—two robotics releases from Apple in a single month.

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, along with 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 typically depend on large volumes 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. That contrast frames the paper's stakes: if general-purpose models can match specialist policies without task-specific training, the large volumes of robot-specific training data behind VLAs and WAMs may not be a prerequisite for capable manipulation.

What It Means

The results come with caveats noted in the reporting: AMP1 relies on third-party models such as Gemini 4 Argon and Claude Opus 5.5, so its performance depends in part on systems Apple does not control. On the other hand, the public release of the code and simulation data means other research labs can test these claims for themselves—making independent replication the clearest next test for the paper's reported numbers.

This article was originally reported by CryptoBriefing and edited by Diego Almada Lopez.