Documentation / Real-robot research

ManiMux¶

An extensible real-world manipulation harness.

Any embodiment. Any policy. Any inference strategy.

ManiMux standardizes deployment and experiment workflows across embodiments, policies, and inference strategies. Built around RoboGUI, it brings together live digital twin visualization and a runtime designed for 100–200 Hz command execution.

Get started Development protocols GitHub

RoboGUI · YAM deployment. Demo footage; some controls have since changed.

One workflow, interchangeable components¶

Model frameworks own inference. ManiMux connects observations, action conversion, chunk scheduling and command execution. RoboGUI lets you operate the experiment; recording and replay let you inspect what happened. Evaluation is optional.

flowchart LR
    O["Observe<br/>Cameras + robot state"] --> P["Predict<br/>Policy framework"]
    P --> T["Adapt + schedule<br/>Adapter → Timeline"]
    T --> E["Execute<br/>Executor → Robot"]
    E --> O
    E -.-> G["Operate + review<br/>RoboGUI · records · replay"]
    style P fill:#fff,color:#233d36,stroke:#377b68
    style T fill:#fff,color:#233d36,stroke:#377b68
    style E fill:#fff,color:#233d36,stroke:#377b68
    style O fill:#fff,color:#233d36,stroke:#377b68
    style G fill:#fff,color:#233d36,stroke:#8d9894,stroke-dasharray:4 3

Architecture and ownership →

Use ManiMux¶

Choose an experiment config, start the services, and work from RoboGUI.

Extend with your agent¶

Give your agent the development skill. It routes each integration to its existing interface and protocol.

Model training and demonstration collection remain in their own tools. Choose your own research questions, experiment plans and evaluation criteria.

Citation¶

ManiMux: An Extensible Real-World Manipulation Harness

Find ready-to-copy BibTeX in the citation guide, including XPolicyLab, StarVLA and PRM-as-a-Judge. Cite the components used in your research.