- Open-RAIL links model inference, robot execution, recording and evaluation.
- The current repository lists three physical robot adapters plus a simulation backend.
- Live teleoperation intervention is on the October roadmap, not part of the current release.
The engineering between a model and a moving robot
China Mobile's TAO team released Open-RAIL on September 16 to connect vision-language-action models with robot hardware. The project packages inference, execution, recording and evaluation into a common software path. It develops work previously published under the name VLA-RAIL.
The problem is timing. A model may need a substantial fraction of a second to produce a batch of future actions, while the robot's control loop needs frequent updates. Waiting for each new prediction can cause pauses; switching abruptly to a new prediction can produce jerky motion. Better reasoning inside a model does not automatically solve that execution problem.
How the earlier research handles continuity
The VLA-RAIL paper describes asynchronous inference and control, with two complementary mechanisms. One smooths the trajectory inside an action chunk using polynomial fitting. The other aligns a newly received chunk with the motion already being executed, preserving continuity in position, velocity and acceleration.
The paper reports tests in simulation and real manipulation tasks. Those results concern the proposed execution method, rather than proof that every supported model can perform every physical task. The September release turns that research direction into a broader engineering framework.
What is available in the repository
The current README lists ten supported models across seven families, including ACT, GR00T, RDT, SmolVLA, GO1, the pi family and TAO. Its robot table names AgiBot G1, China Mobile Lingxi and NAVIAI-WA2, plus a LeRobot-based simulation backend. This is more precise than describing the list as four physical robots.
The server runs the model environment and the client handles the robot. That separation helps keep model dependencies apart from robot-driver and ROS dependencies. The framework records observations and actions in a LeRobot-style format and writes timing logs alongside the run configuration. The project is Apache-2.0 licensed.
A useful boundary in the roadmap
The repository marks live teleoperation intervention and certain additional integrations as October work. A planned pause-intervene-resume workflow should therefore not be described as already available in the September version. Likewise, a mock backend can check that software components communicate; it cannot demonstrate real task performance.
The architecture aims to make an experiment more reusable. Researchers can exchange a model or adapt a robot without rebuilding every surrounding component. Recorded execution data can then help explain whether a failed task came from perception, inference delay, communications or motion handling.
What researchers can inspect after a run
Open-RAIL records inference delay, image-preprocessing time, communication delay, observation frame rate and the processing time within and between action chunks. It exports raw timing arrays and CSV averages on a common time base. The accompanying configuration records the model, control period and smoothing choices, so a comparison can separate a slower model from a slower link to the robot.
The logs include timeouts, missing observations and manual aborts, not just completed trials. The repository also provides a worked GR00T-N1.5 example using the AgiBotWorld 2026 dataset. Together, these artifacts give a new user a concrete starting point for testing the software without interpreting a polished demonstration video as the full experimental record.
Why this matters beyond a demonstration
Robotics progress often appears as a video of a completed action. A reusable deployment layer adds something different: the machinery to run another trial, record what happened and compare configurations. That is the practical significance of Open-RAIL's release.
It remains an integration framework, rather than a universal robot skill or an assurance of safe autonomous operation. The model, physical platform and task constraints still determine what the complete system can do.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01Updates & corrections
— Expanded using the official repository and paper. Clarified that the repository's four listed backends include a simulation backend, and that teleoperation intervention is planned for October.