How China Mobile’s Open-RAIL Connects AI Models to Robots

An AI model can plan a robot’s next movements, but it does not necessarily produce commands at the rate the robot needs them. China Mobile’s open-source Open-RAIL sits between vision-language-action (VLA) and world-action (WAM) models and robot hardware to address that practical gap. It is an execution layer, not a new model: its job is to keep control running, smooth model-generated actions and provide interfaces for different machines. 69

Keeping the robot moving between predictions

Open-RAIL separates model inference on a server from sensing and execution on a robot-side client. Observation, inference and control run asynchronously, allowing the robot to execute one action chunk—a batch of planned commands—while the model computes the next. Release reports put model inference at roughly 5–10 Hz, or predictions per second, and robot control at 200–500 Hz. 36

Continuous control alone does not ensure smooth motion. Open-RAIL first smooths the trajectory within an action chunk, then blends a newly returned chunk with the movement already underway. The earlier VLA-RAIL preprint describes the goal as continuity in position, velocity and acceleration across chunk boundaries. This processing happens in the execution layer without changing the model’s weights. 123

Swapping models and machines

On the hardware side, a lightweight abstraction layer gives upper-level software common ways to read robot state and issue actions; a new machine can be connected through a driver adapter. On the model side, a unified inference interface lets registered model adapters supply actions through the same workflow, with routing connecting the selected components. That can reduce integration work, but an adapter alone does not make every model’s actions suitable for every robot. 3810

China Mobile reports four robot backends and ten model entries in the first release. Descriptions of the support table identify three physical platforms—Agibot G1, Ti5 T170C and Navi WA2—plus a LeRobot-based simulation backend. The listed model entries span ACT, GR00T N1 variants, RDT-1B, SmolVLA, AgiBot GO-1, π0 variants and TAO. The published totals are not a tested, all-to-all model–robot compatibility matrix. 918

The reported gains—and the limits of the evidence

China Mobile says joint-acceleration standard deviation fell from more than 10 to about 0.1 rad/s². It also says connecting a new robot can take hours rather than weeks, while a new model can require roughly 50–100 lines of adapter code. These are project-reported figures, not independent benchmarks across all supported setups. 611

Open-RAIL archives observations, robot states and inference results during real executions for later training and evaluation. Reports say it can also save human corrections aligned with the model’s actions when an operator intervenes. Its server–client design allows inference to run on a device, at the edge or in the cloud without changing upper-level application code, according to release coverage. The cited results do not establish the collection overhead or equivalent performance in all three locations. 5611

The Open-RAIL project page links to its code and documentation; release coverage describes Apache-2.0-licensed code available through GitHub and Gitee. For the technical background on asynchronous execution and smoothing, see the earlier VLA-RAIL preprint. 139