Apple researchers Chen Chen and Omid Bagherieh have published AMP1, short for Agentic Manipulation Policies, a framework that lets general-purpose AI models control robot arms by writing and executing Python code rather than relying on robot-specific training data. In 108 real-world trials on Franka robot arms, AMP1 completed 104 successfully, a 96.3% success rate, with zero fine-tuning or demonstration examples for the specific tasks tested.
The paper, titled "On Agentic Manipulation Policies" and posted to arXiv in November 2026, pairs AMP1 with existing large language models including Gemini 4 Argon and Claude Opus 5.5, rather than building a new foundation model from scratch. Instead, AMP1 gives those general-purpose models a code-writing interface that can issue motion commands, call tools, and search for solutions in real time, letting the AI reason about a manipulation task the same way it would reason about writing software.
Beating Specialist Robot Models at Their Own Game
The result challenges an assumption that has shaped robotics research for years: that robot manipulation needs purpose-built, heavily trained policies learned from thousands of demonstrations. In simulation benchmarks, AMP1 paired with Gemini 4 Argon scored 85.6% on RoboLab-120 and with Claude Opus 5.5 scored 82.8%, both ahead of Nvidia's Astranomy baseline at 74.8%. On LIBERO-PRO and Robosuite, two more established manipulation benchmarks, AMP1 reached 95.2% and 100% respectively. In the real-world arm trials, baseline systems the researchers compared against completed only 49 and 30 successful trials out of equivalent attempts, meaning AMP1 cut operational failures by roughly 15 to 20 times.
Those numbers put AMP1 in direct comparison with specialist robot foundation models from companies racing to build general-purpose robot "brains," including Nvidia's GR00T line, Google DeepMind's Gemini Robotics models, and physical AI models from startups such as Physical Intelligence and Skild AI. Most of those systems are trained specifically on robot demonstration data, often collected through teleoperation or simulation at significant cost. AMP1's pitch is that a model never trained on robotics data at all, when given the right code-execution scaffolding, can match or beat those purpose-built systems on standard tasks.
What It Means Going Forward
Apple has made AMP1's code and simulation data publicly available, a departure from the company's traditionally closed approach to research releases and one that invites outside labs to replicate or extend the results. If the approach holds up outside Apple's own test rigs, it suggests robotics companies could lean more on general-purpose AI providers for the reasoning layer of manipulation tasks, reserving specialist training for narrower edge cases that code-writing alone can't solve. The next test will be whether AMP1 performs as well on messier, less structured tasks outside controlled lab trials, and whether its reliance on third-party models from Google and Anthropic creates dependency risks Apple would rather avoid in its own robotics ambitions.