Teaching a robot to “put the bowl on the plate and place the cream cheese inside the bowl” used to require a rack of data-center GPUs sitting behind the robot. In this blog we show how a modern vision-language-action (VLA) model, MolmoAct2 , can look at a scene, read a plain-English instruction, and drive a robot arm to complete the task, all running on a single AMD Ryzen AI Max+ 395 “Strix Halo” mini PC with the ROCm software stack. No cloud, no external accelerator, just a compact tabletop machine. By following this blog, you will learn how to deploy and run a state-of-the-art VLA model optimized for the AMD Ryzen AI Max+ 395. You will use a reproducible Ryzers container to launch browser-based robot control, test zero-shot transfer across robot arms, run asynchronous real-time planning, and evaluate the policy in simulation and against real-world data. These workflows give you a practical starting point for prototyping and benchmarking robotics applications locally with ROCm, without relying on cloud or data-center GPUs. We package everything as a ready-to-run Ryzers container so anyone can see the robot in action with a few commands. Along the way we demonstrate that the robot follows instructions inside the LIBERO environment, show that the same policy transfers to robot arms it was never trained on, run it in real time while it plans and moves at the same time, and validate it against real-world robot data. As you can see in the video figure (Figure 1) below, the policy drives the arm directly from a browser session served by the mini PC. Figure 1. Controlling a Panda robot arm in the LIBERO simulator, straight from a web browser, with MolmoAct2 running on a Ryzen AI Max+ 395 Mini PC. What is a Vision-Language-Action model? Traditional robots are programmed with hand-written rules on top of task-specific closed-loop controllers. A vision-language-action model takes a very different approach: it is a foundational neural network that consumes what the robot sees (camera images) and what you ask it to do (a natural-language instruction) and directly produces what the robot should do next (motor commands). It combines three ideas that have each matured enormously in recent years—computer vision, language understanding, and robot control—into one unified model. In a previous post we showed how to bring the LeRobot robot learning framework to ROCm. Here we show the deployment of a full foundational VLA model. MolmoAct2 is a particularly exciting milestone for open robotics: it offers fully open source weights, training code, and data. It also generalizes across multiple robot embodiments from a single trained model, rather than one model per robot. It also builds on Ai2’s open Molmo lineage (including the MolmoSpaces environments) for grounded 3D understanding. MolmoAct2 is an open, state-of-the-art VLA family. It is based on the MolmoAct2-ER model which, under the hood, pairs a SigLip2 vision encoder and a Qwen3-4B language backbone with a flow-matching action expert that generates short “chunks” of future motion. The policy speaks in end-effector coordinates – it decides where the robot’s hand should go and whether the gripper should open or close. As we will see, that design choice unlocks a pleasant surprise: cross-embodiment policy transfer. It can also command individual joint motors directly. The open source nature of the model and the dataset makes MolmoAct2 a very convenient tool to play around with. VLA models are large, but the Ryzen AI Max+ 395 pairs a capable integrated GPU with a large pool of unified memory, so it can hold and run these models comfortably on a device that fits on a desk. Tutorial: A Single Docker Image for Commanding The Robot Army The MolmoAct2 Ryzers package bundles a complete set of demonstrations, each of which exercises a different part of the “see, think, act” loop: Interactive simulation : command a Panda arm from the browser in the MuJoCo environment. Cross-embodiment : run the same poli