China Mobile releases Open-RAIL humanoid control software
China Mobile has released Open-RAIL, an open source deployment layer designed to reduce jerky motion when vision language action models control humanoid robots. The software separates observation, model inference and motor control into concurrent processes, then smooths the resulting movement commands.
According to Forbes, current vision language action models typically generate short chunks of planned actions five to 10 times per second, while robot motors can require updated commands at rates between 200 Hz and 1 kHz. The mismatch can produce stalls or abrupt movement at the boundaries between action chunks. The Open-RAIL researchers say some demonstrations compensate by slowing execution by a factor of four to eight.
Separate control loops and two smoothing stages
Open-RAIL runs observation at 30 Hz, model inference at five to 10 Hz and motor control at 200 Hz to 1 kHz. While the robot executes one action chunk, the model can calculate the next rather than leaving the body waiting for another command.
A two stage process first smooths motion within each chunk and then blends the boundaries between consecutive chunks. The arrangement loosely resembles the division of labor in human motor control, where higher level planning operates alongside faster local corrections.
The researchers report that Open-RAIL reduced jerkiness by 100 times, more than doubled execution speed and improved task success rates by 10 to 73 percent. Those results have not been independently replicated, and Forbes noted that the project site does not provide a task by task breakdown. The size of the claimed gains therefore needs to be read as a result reported by the development team rather than an established benchmark.
Deployment tooling beyond motion smoothing
Open-RAIL humanoid control is presented as more than a smoothing algorithm. It includes a server client architecture, a web control panel with more than 40 runtime parameters, automatic recording in Hugging Face LeRobot’s data format and human in the loop teleoperation. Operator corrections can be captured and returned to the training dataset.
The team says integrating another model requires 50 to 100 lines of code. Its hardware abstraction layer is intended to reduce the time needed to add another robot from weeks to hours, although that onboarding claim remains theoretical.
Supported models include Nvidia GR00T, Physical Intelligence’s π0 and π0.5, ACT, RDT, SmolVLA, AgiBot GO-1 and China Mobile’s TAO. Listed robot platforms include the AgiBot G1, NAVIAI WA2 and China Mobile Lingxi, while the project site also names the Unitree G1.
Open-RAIL is not the only attempt to handle action chunk boundaries. Physical Intelligence published a real time chunking method in 2025, and Hugging Face incorporated it into LeRobot. The Open-RAIL paper compares its approach with that method as well as A2C2 and VLASH. Its open release should make independent testing across different humanoid hardware and manipulation tasks possible.
Source: Forbes
