molmoact2

MolmoAct2

This package runs MolmoAct2 VLA model on AMD Ryzen AI Max+ 395 (Strix Halo Mini PC). Demos include:

• A browser-based interactive LIBERO simulator with the Panda robot arm.
• Zero-shot cross-embodiment interactive demos for UR5e and xArm6 robot arms.

Build

ryzers build molmoact2
ryzers run

ryzers run with no command is a light ROCm environment check.

Artifacts are written to workspace/molmoact2/outputs.

For downloading weights faster from huggingface, set the environment variable HF_TOKEN to your own access token: export HF_TOKEN=XXXX....XXXX

Interactive Demo

Interactive browser-based control of a Panda robot in the LIBERO MuJoCo environment.

To command a robot arm to perform tasks in the LIBERO-MuJoCo simulation environment from a browser:

ryzers run /ryzers/demo_interactive.sh

Then open in a browser:

http://localhost:8080

The synchronous demo defaults to the fast path and skips depth reasoning. For the full depth reasoning model, pass THINK=1.

Complex Task Example

Prompt: "Put bowl on plate and put cream cheese inside bowl"

Cross-Embodiment Interactive Demos

The policy action operates in end-effector coordinates while a robot controller handles joint movements using inverse kinematics. As a result, the MolmoAct2 policy transfers naturally across robot embodiments in simulation.

UR5e with Robotiq85 Gripper

EMBODIMENT=ur5e ryzers run /ryzers/demo_interactive.sh

Zero-shot transfer of MolmoAct2 to a UR5e robot equipped with a Robotiq85 gripper.

xArm6

EMBODIMENT=xarm6 ryzers run /ryzers/demo_interactive.sh

Zero-shot transfer of MolmoAct2 to an xArm6 robot embodiment.

Real-Time Interactive Option

The real-time server runs the simulator at wall-clock speed while policy planning occurs asynchronously. The robot holds its current pose while the next action chunk is computed.

ryzers run /ryzers/demo_interactive_rt.sh #optionally pass EMBODIMENT=xarm6 or EMBODIMENT=ur5e 

To see the demo in a browser, open:

http://localhost:8081

For longer-horizon tasks (raise the max episode steps; default 1200):

RT_MAX_STEPS=3000 ryzers run /ryzers/demo_interactive_rt.sh

Closed-Loop LIBERO Evaluation

Run a closed-loop LIBERO rollout in MuJoCo:

ryzers run /ryzers/demo_libero.sh
SUITE=libero_object TASK_ID=3 ryzers run /ryzers/demo_libero.sh

Available suites:

• libero_10
• libero_goal
• libero_object
• libero_spatial

TASK_ID ranges from 0 to 9.

Closed-loop LIBERO evaluation rollout in MuJoCo.

DROID Open-Loop Replay

Run MolmoAct2-DROID open-loop on a dataset episode:

ryzers run /ryzers/demo_droid.sh
EPISODE=42 ryzers run /ryzers/demo_droid.sh

The demo generates a scene video and a ground-truth versus predicted action plot.

Open-loop replay on a DROID dataset episode.

Useful Knobs

• THINK=1 enables the full depth reasoning path.
• THINK=0 NUM_STEPS=4 is the fast interactive default.
• PORT=... changes the browser port.
• SEED=..., SUITE=..., and TASK_ID=... make runs reproducible.
• ryzers run /ryzers/download.sh preloads the MolmoAct2-DROID, MolmoAct2-DROID-Dataset, and MolmoAct2-Think-LIBERO assets into the mounted cache.

Copyright (C) 2026 Advanced Micro Devices, Inc. All rights reserved.