Robots in general and humanoid robots specifically can look a little like really old movies or badly buffering YouTube videos: shaky and jerky. China Mobile just open-sourced a fix. It’s called Open-RAIL , and it solves a mismatch between robot brains, or AI models, and robot bodies.

The AI models that run robots, typically vision-language-action models (VLAs) like Nvidia’s GR00T and Physical Intelligence’s π0, can be a bit slow. They look at the world, think, and output an “action chunk," or a short burst of planned movements, about five to 10 times a second. Robot motors, however, want fresh commands 200 to 1,000 times a second.

That’s a gap of 30X or more.

“Existing methods suffer from jitter, stalling, or even pauses in robotic action execution, which not only limits the achievable execution speed but also reduces the overall success rate of task completion,” the China Mobile researchers write in their paper .

Five to 10 times per second sounds fast, and it is, in a sense. (In fact, our brains work at similar speed .) But robot CPUs and GPUs are running the robot body, integrating data from multiple sensors including video, LIDAR, and audio, connecting to a central management system, and many other things. Also, VLAs are huge neural networks, and calculating anything takes time, even with advanced chips and plenty of memory. So when their bodies start executing commands, then wait for the next one, then move, then wait, the result is jerky, awkward, and wasteful movement.

In fact, if the robot is pouring tea or coffee, it’s also potentially messy if not dangerous.

The standard response right now? Just go slow.

According to the paper, many VLA demonstrations slow execution down by a factor of 4 to 8 to hide the jitter. Which might explain why some humanoid robot demo videos come with a little “2X” or “4X” in the corner. And which might explain why humanoids tend to be a little slower than you’d expect.

The new open source software works by ensuring the robot body doesn’t need to wait for new commands from the robot brain. It splits movement into three independent threads: observation at 30 Hz, or times per second, model inference at 5-10 Hz, and motor control at 200 Hz to 1 kHz. That means that while the robot executes one chunk, the model is already working on the next. Then a two-stage smoothing routine cleans up the result: first smoothing the motion inside each chunk, then blending the seams between the chunks of motion.

Result: clean, simple, smooth motion … perhaps even almost human-like.

This is roughly analogous to what our own human brains and bodies do. Muscles help by being soft and springy, smoothing out motion, of course, but our spinal cord handles fast corrections that don’t even need to hit the brain, and our brain plans ahead, predicting the future and prepping future motions before current ones end. Movements also overlap – like the two-stage routine blending the seams between discrete motions in Open-RAIL – with one starting just before the previous one ends, like reaching for a cup and lifting it before we’ve even fully completed the grasping part of the motion. Our brain also aims for smoothness , like Open-RAIL’s smoothing.

The results look promising.

According to the researchers, the software cut jerkiness by 100X. That’s kind of the difference between a kid learning to drive using stick and a high-end automatic transmission. Open-RAIL can also more than double robot speed and boosts task success rates from 10-73%. Those are the team’s own numbers, of course, and the project site doesn’t break them down task by task. Independent replication would help … and should come soon, thanks to the fact that this software is now open source and freely available.

Others have different methods of solving this problem, of course. Physical Intelligence published its own real-time chunking method in 2025, and Hugging Face has built it into its open LeRobot framework. The Open-RAIL paper benchmarks itself against that approach as well as other academic methods like A2C2 and VLASH.

Interestingly, however, Open-RAIL is not so much a single algorithm as a full deployment layer including a server-client system with a web control panel exposing 40-plus runtime parameters. It also does automatic data recording in LeRobot’s format and human-in-the-loop teleoperation so an operator can correct a robot mid-task and feed that correction back into training data.

Adding a new model takes 50 to 100 lines of code, the team says, and a standard hardware abstraction layer will theoretically cut new-robot onboarding “from weeks to hours.”

So far, supported models include Nvidia GR00T, Physical Intelligence’s π0 (spoken as “pi-zero”) and π0.5, ACT, RDT, SmolVLA, AgiBot’s GO-1, and China Mobile’s own TAO model.

Supported robots include the AgiBot G1, NAVIAI’s WA2, and China Mobile’s Lingxi, with the project site also listing Unitree’s G1.