AMD Runs MolmoAct2 Robot AI Locally on Ryzen AI Max+ 395: ROCm Setup and Simulation Limits

What AMD actually built​

The pitch is that no cloud and no datacenter GPU are needed, only unified memory, Ryzers containers and zero-shot cross-embodiment control. Ryzers is AMD's lightweight framework for building and running ROCm Docker containers on Ryzen AI hardware. The AMDResearch repository lists MolmoAct2 among its VLA packages, next to OpenVLA, SmolVLA and GR00T-N1.5. It describes the project as composable Dockerfiles for Ryzen AI machines. It also notes minimal host requirements, with standard Ubuntu support.

The model pairs a SigLip2 vision encoder and a Qwen3-4B language backbone with a flow-matching action expert. The action expert produces short chunks of future motion. The policy works in end-effector terms. It decides where the hand goes and whether the gripper opens or closes. It can also command joints directly.

The five demonstrations​

The container bundles five workflows:

• Interactive control. A Panda arm in a LIBERO scene (MuJoCo) is driven from a browser. The headline task is "Put bowl on plate and put cream cheese inside bowl."
• Cross-embodiment. The same policy is pointed at a UR5e with a Robotiq 85 gripper, or at an xArm6.
• Real-time mode. The simulation runs at wall-clock speed while the policy plans the next action chunk in the background. The arm holds its pose until the new plan arrives.
• Closed-loop evaluation. Four LIBERO suites (libero_10, libero_goal, libero_object, libero_spatial) each have task IDs 0 to 9.
• Open-loop replay. A recorded DROID episode is replayed. The model's predicted actions are plotted against what a human teleoperator actually did.

Reproducing it​

The commands below come from AMD's tutorial.

1. Install the framework. Run git clone [GitHub - AMDResearch/Ryzers: Ryzen AI Dockerfiles for ML, Robotics and Development Environments · GitHub](https://github.com/AMDResearch/Ryzers), then pip install Ryzers/.
2. Build the package with ryzers build molmoact2.
3. Run ryzers run with no command. This performs a light ROCm sanity check.
4. Optionally run export HF_TOKEN=... first. A Hugging Face token speeds up the first-run download of weights and datasets. Those are cached and reused afterwards.
5. Launch the browser demo with ryzers run /ryzers/demo_interactive.sh. Then open [url]http://localhost:8080[/url].

The other demos use the same pattern:

The synchronous demo skips the optional depth reasoning by default, which keeps it responsive. Setting THINK=1 enables it. Other variables include PORT, SEED, SUITE, TASK_ID and NUM_STEPS. Videos and plots are written to a local outputs folder.

Reading the claims carefully​

Several claims need qualifying, and some of that comes from Ai2's own repository.

Zero-shot transfer is a simulation demo. Ai2's repository says out-of-the-box deployment targets simple tasks within the training distribution, such as pick-and-place, opening and closing. It says performance has been empirically verified only on SO-100 and Franka DROID. It lists SO-100, Bimanual YAM and Franka DROID as the supported setups. UR5e and xArm6 are not on that list. AMD's cross-arm result is therefore a demonstration in AMD's package. It is not Ai2-validated hardware support.

The tutorial has no numbers. The article gives no inference latency, no task success rates, no power draw and no comparison with data-center GPUs. "Comfortably" is a qualitative word. The article also doesn't say how much memory the test mini PC had. AMD lists up to 128 GB as the platform maximum, but that isn't the tested configuration.

Open-loop replay is not a live trial. The replay compares predictions against recorded human actions. AMD says the predictions "closely matched" the demonstrations, but it publishes no metric. A model can track a recording well and still fail when its own errors compound in a live loop. That is the reason to run closed-loop tests too.

Asynchronous planning is not seamless control. The arm freezes while a plan is computed. The authors describe removing that pause as ongoing research.

Why it matters for PC users​

The model is open, so it isn't tied to one vendor. Ai2's paper says MolmoAct2 is fully open and releases weights, training code and data. Its abstract complains that open-weight alternatives are tied to expensive hardware. An unverified-but-reproducible path to run it on a desktop-class integrated GPU speaks to that complaint. The Ryzen AI Max+ 395 is the hardware that has made local AI practical for hobbyists.

Community reports show Strix Halo's ROCm story has been uneven. One LM Studio issue describes gfx1151 problems that went away with ROCm 7.2 or later. A community benchmark thread found ROCm 6.4.4 much faster than 7.0.1 on one test. Treat the container as the supported route, and expect to adjust your ROCm setup if you go off it.

Safety​

Ai2's guidance says to validate outputs before using a physical robot. It recommends simulation rollouts, action limits, emergency stops and human supervision. For anyone attaching real hardware, run the closed-loop LIBERO tasks first. Then add speed, workspace and force limits before attempting any physical test.

Bottom line​

The package makes a large open robotics model easy to try on a Ryzen AI Max+ 395 box with a few commands. It shows that local inference works end to end in simulation and on recorded data. It doesn't show that the same setup is reliable on physical robots, and the sources contain no benchmarks to support that.