AMP1 lets general-purpose AI models control robot arms with code

AMP1 stands for Agentic Manipulation Policies.[1] The paper describes it as a reference harness that runs unchanged across single-arm, bimanual and mobile manipulation.[2] Rather than relying on demonstrations or task-specific training, AMP1 is designed to operate without either.[2] Researchers also say it can adjust its actions as a task proceeds.[2] For teams evaluating robot-control approaches, the central proposition is that a general-purpose model can be used across different manipulation setups through a common harness, rather than a separately trained system for each task.[1][2]

The basic AMP1 harness does not supply prebuilt grasping, pushing or placing skills.[2] Yet the paper says the system can develop strategies that include searching for hidden objects, using objects as tools and calibrating cameras.[2] That distinction matters when interpreting the results: the reported approach is not described as a library of ready-made manipulation behaviors, but as a harness that lets a model produce and adapt actions.[1][2] The paper does not establish that every model or task will produce the same strategies or outcomes.

The project page reports AMP1 results of 95.2% on LIBERO-PRO, described as its best zero-shot result on that benchmark, and 100% on robosuite.[3] It also reports 96.0% on BEHAVIOR-1K for mobile manipulation.[3] On RoboDojo, the page gives an 84.1% result and describes it as the best public number after practice.[3] Separately, the researchers report that an augmented AMP1 harness lowered failure rates on RoboDojo bimanual tasks by 4.6 times compared with zero-shot deployment.[2] These are benchmark-specific results, not a single measure of performance across all robots or tasks.

In real-world tests, AMP1 succeeded in 104 of 108 trials spanning 12 tasks and three robot instances.[2] The researchers also report 20-fold lower failure rates against a strong vision-language-action baseline and 15-fold lower rates against a strong world-action-model baseline.[2] The evidence describes those comparisons as failure-rate reductions; it does not provide a basis for treating them as proof that AMP1 outperforms every specialist system. The paper was submitted to arXiv on November 6, 2026.[2] Teams considering the results should distinguish benchmark scores, real-world trials and baseline comparisons.