AI · September 26, 2026 · 2 min
China Mobile open-sources Open-RAIL to steady jerky humanoids
The layer lets a slow vision-language model plan while a faster loop runs the motors. The team says it cut jerk by about 100 times.
China Mobile released Open-RAIL, an open-source layer that sits between a humanoid’s AI model and its motors. Forbes and Humanoid.guide described it as a fix for the mismatch that makes vision-language-action robots look jerky: the model thinks slowly, and the body should not wait on every new command.
Open-RAIL watches the scene at 30 hertz, runs the model at 5 to 10 hertz, and sends motor commands at 200 hertz to 1 kilohertz. While the robot plays one chunk of motion, the model can compute the next. A two-step smoother cleans motion inside a chunk, then blends the seam into the following one.
The researchers say that cut jerkiness by about 100 times, more than doubled execution speed, and lifted task success by 10 to 73 percent. Those figures are the team’s own. Supported models include Nvidia GR00T, Physical Intelligence’s π0 and π0.5, and AgiBot GO-1. Named robots include the Unitree G1 and the AgiBot G1.