Open-RAIL connects VLA models and humanoids, smoothing and speeding motion
Open-RAIL smooths and speeds humanoid robot motions
5-10 Hz to 200-1,000 Hz.
That's the mismatch every team deploying a vision-language-action model or world action model (VLA or WAM) on a humanoid runs into: big models plan a few times a second, but motors need commands hundreds of times a second. China Mobile's embodied AI team thinks it has a general-purpose fix, and on September 16 it released it as open source under Apache 2.0.
Open-RAIL (Real-time Asynchronous Inference Linker) is middleware. It doesn't make a robot smarter. It makes a robot execute what its model already decided without the stutter, stalls and pauses that come from stitching one action chunk to the next.
Its origin is a December 2025 paper from China Mobile (Hangzhou) Information Technology, where it was called VLA-RAIL.
How it works
The architecture is straightforward: three asynchronous threads for observation (30 Hz), inference (5-10 Hz) and control (200 Hz-1 kHz), so the robot keeps moving while the model computes its next chunk. Then two smoothing passes. Inside each chunk, polynomial fitting removes noise. Between chunks, quintic spline fusion blends the old plan into the new one so there's no discontinuity when fresh actions arrive.
The claimed gains, from the team's own testing:
- Joint acceleration standard deviation drops from 10+ rad/s² to about 0.1.
- Execution runs up to 2.09X faster than raw policy output.
- Task success reaches as high as 95%, with absolute gains of 10 to 72.5 percentage points, tested on π0.5 and GR00T-N1.5.
The paper also notes the workaround most teams use today: "many VLA demonstrations slow down the execution of action chunks to reduce the impact of jitter, sometimes by a factor of 4 to 8."
That's the cost Open-RAIL is going after.
The choice of quintic splines (smooth curves that match up beginning and ending position, speed, and acceleration) is interesting. Fifth-order polynomials are the same form as the classic minimum-jerk model of human arm movement, which Tamar Flash and Neville Hogan published in 1985. Whether by design or not, the smoothing math borrows from the shape that has long been used to describe how people move.
What matters more for the industry, though, is the scope. This is closer to a deployment operating system than a smoothing trick.
The server hosts the model. The client handles perception, trajectory smoothing, execution and teleoperation. A web panel exposes 40+ runtime parameters and records every episode in LeRobot-style Parquet, so real-world runs turn directly into training data. Operators can take over mid-task via human-in-the-loop correction, and those corrections get logged too. China Mobile calls this "closed-loop evolution." Integrating a new model reportedly takes 50-100 lines of code.
Supported models and robots
Model support spans much of the open VLA landscape: Nvidia GR00T N1 through N1.7 (per the project page), Physical Intelligence π0 and π0.5, ACT, RDT-1B, SmolVLA, AgiBot GO-1, DreamZero and China Mobile's own TAO.
Robot support is where it gets interesting for anyone tracking the humanoid market. The adapted platforms are AgiBot G1, NAVIAI WA2, China Mobile's Lingxi (built on Ti5's T170C) and, according to the project site, Unitree G1. Three of those four are wheeled, dual-arm platforms. Open-RAIL is focused on upper-body manipulation, not bipedal locomotion — that remains a separate control problem handled by each robot's own low-level controllers.
A note on the numbers: the project page says 20+ supported models and lists Unitree G1, while the GitHub README says 10 models across 7 families and lists a LeRobot simulation backend as the fourth "robot." Early-release docs often drift, but builders should check the repo for what's actually implemented today.
Of course, there are other approaches
The competition here is mostly research code. Physical Intelligence's real-time chunking (RTC) is the best-known approach and is already integrated into Hugging Face LeRobot. The VLA-RAIL paper benchmarks against RTC plus A2C2 and VLASH. Open-RAIL's pitch is that it wraps an approach like that in a full, model-agnostic, robot-agnostic toolchain.
Then there's the carrier angle. Open-RAIL supports three deployment tiers with published latencies: on-robot (3-5 ms), edge (35-45 ms) and cloud (80-120 ms). A 35-45 ms edge round-trip is well within a 5-10 Hz inference budget of 100-200 ms per cycle. For a telecom operator, that's a case for robot inference as a network service.
It also continues a clear pattern: Chinese players open-sourcing the humanoid stack, from Beijing Humanoid's XR-1 VLA to the full Tien Kung robot, just as U.S. regulators move to block Chinese robot hardware.
Software travels more easily.
Adoption, as usual, is the open question. At launch the repo had 33 stars and 11 forks.
Sources:
Written by
I tell stories with data and build startups. I've raised angel and VC funds, led software companies and startups, and chronicled the rise of the mobile economy. Currently I write for Forbes, host several podcasts, and am starting up HumanoidMegaHub. I was TUNE's mobile economist and Singular's VP of Insights, built the research team at VentureBeat, and managed teams creating software for partners like Intel and Disney. In past lives I built websites, e-learning platforms, and mobile apps, and, a long time ago, was named to Folio's top 100 of the media industry's "most innovative entrepreneurs and market shaker-uppers.”