How does China Mobile’s open source Open RAIL middleware connect VLA and WAM models to heterogeneous robot hardware, and what problems does it address through its asynchronous server–client pipeline, two level action smoothing, hardware abstraction, unified inference interface, a

Open-RAIL is China Mobile’s middleware between vision-language-action (VLA) or world-action (WAM) models and robot bodies. It is intended to make slow, chunk-based model inference work with continuous robot control, while allowing models and hardware to be changed without rebuilding the whole execution stack. 37

How it works

• Asynchronous server–client pipeline: The server handles model inference, scheduling and data management; the robot-side client handles sensing and execution. Separate observation, inference and control flows let control continue while the model computes its next action chunk, addressing pauses caused by slower inference. 311
• Two-level action smoothing: One stage smooths points within an action chunk; another fuses a newly returned chunk with the trajectory already being executed. This targets both jitter inside a chunk and abrupt motion at chunk boundaries without changing model weights. 211
• Reusable interfaces: A lightweight hardware-abstraction layer standardizes state reads and action execution, so a new robot needs a driver adapter rather than a rewrite of upper-level logic. Models implement a common inference interface; automatic routing selects the registered model and robot adapters, reducing model–hardware integration work. 37

Reported results and first-release scope

• China Mobile reports that joint-acceleration standard deviation fell from more than 10 rad/s² to about 0.1 rad/s², and that onboarding a new robot fell from weeks to hours; a new model can reportedly require only 50–100 lines of adapter code. These are reported figures, not independently established results. 117
• The release is reported to cover four robot backends and ten model variants: Agibot G1, Ti5 T170C, Navi WA2 and a LeRobot-based simulation backend; and the ACT, GR00T N1/N1.5/N1.6, RDT-1B, SmolVLA, AgiBot GO-1, π0/π0.5 and TAO model families. The available evidence does not establish that every model–robot pairing is supported. 9

Execution-data collection is integrated into the inference-to-execution workflow: real runs accumulate training-ready data for later model iteration, rather than requiring a separate collection process. The server–client split also permits inference to be placed on device, at the edge or in the cloud without changing the robot-side application code; the evidence here does not specify every recorded field or an independently measured collection overhead. 5814

The Apache-2.0 code and documentation are linked from the Open-RAIL project page; reports describe GitHub and Gitee copies, but the evidence available here does not establish their exact repository URLs. The earlier VLA-RAIL preprint describes the asynchronous linker, trajectory smoother and chunk fuser. 92