China Mobile Open-Sources Open-RAIL: The Middleware Layer Connecting VLA Models to Robot Hardware

The Problem: Two Clocks Out of Sync

When a robot arm must complete a long sequence—grab a cup, place it, press a button, wait for the drink, retrieve and transfer the cup—there is a fundamental timing conflict: model inference takes on the order of 100 ms per action step, while joint control loops demand a new command every 3.75 ms. That is a 30-50x gap. Under synchronous execution, the lower bound of control-cycle latency is the inference latency, and the physical result is concrete: every time the model "thinks," the robot stops.

On September 16, China Mobile open-sourced Open-RAIL globally, positioning it as the industry's first general-purpose engineering foundation connecting VLA / WAM models with robot hardware. It sits neither on the model side nor the robot side, but in the middleware layer that every team has previously had to rebuild themselves.

Beyond the clock mismatch, three related problems compound the difficulty:

• Interfaces: different robot brands have different joint definitions, communication protocols, and coordinate conventions; switching models or robots means redoing mapping work.
• Seams: models output discrete action points with no smooth transitions between segments, causing sudden jumps in pose and velocity at segment boundaries.
• Data: inference and data collection are traditionally separate pipelines; cleaning formats and aligning timestamps takes weeks.

Architecture: Server-Client Split with Three Parallel Pipelines

Open-RAIL separates a server side (task scheduling, model inference thread pool, real-time data management with timestamp synchronization) from a client side on the robot. This division exists for a practical reason: the deep-learning environment models need and the control environment robot drivers need would otherwise conflict.

On the robot side, observation, inference, and control run as three parallel pipelines that never wait on each other, with an action buffer absorbing the rhythm difference between the two ends. The robot no longer stops when the model thinks.

Asynchronous execution only solves the "waiting" problem. Two further jitter sources remain: sparse sampling inside a single inference result (intra-chunk) and the seam between the currently executing trajectory and the next inference result (inter-chunk). Open-RAIL addresses both levels simultaneously, computed in real time on the robot side, with zero changes to the model.

The most concrete number: joint acceleration standard deviation drops from 10+ rad/s² to 0.1 rad/s²—two orders of magnitude. Smoothing suppresses acceleration jerk, which is where liquid sloshing, structural shock, and motor inrush currents come from.

End-Edge-Cloud: Decoupling Compute from the Robot

Open-source VLAs intended for real robots range from hundreds of millions to billions of parameters, with around 3B the most common—not because larger models aren't wanted, but because embedded devices on robots can't fit them while competing with millisecond-level control loops.

Open-RAIL puts all inference on the server side, which can run on the robot itself, on edge compute, or in the cloud—the code is identical, only the connection address changes. This unlocks the constraint that limited on-board compute forces small models, at the cost of network availability becoming a new prerequisite for cloud deployments.

Heterogeneous Adaptation

A lightweight hardware abstraction layer unifies control, state reading, and action execution behind one convention:

| Item | Before | After | |---|---|---| | New robot adaptation cycle | Weeks | Hours | | New model integration code | Full inference pipeline rewrite | 50-100 lines | | Robots adapted | — | 4 heterogeneous models | | Models supported | — | 10 mainstream VLA / WAM models |

The four adapted robots are China Mobile's Lingxi folding wheeled-arm robot, AgiBot (Zhiyuan) G1, Unitree G1, and Zhejiang Humanoid wa1—machines with different DOF configurations, joint naming, and coordinate systems, all running the same upper-layer logic.

Data Loop: Collection as a Byproduct of Running

Unlike pure inference libraries, Open-RAIL automatically archives multi-view images, joint states, inference outputs, smoothed commands, and timing data during execution, in training-ready standard formats, written asynchronously without blocking the control thread. Per-stage latencies (inference, smoothing, communication) are logged separately, with run configs saved for reproducibility.

A hybrid "inference + teleoperation" mode lets humans intervene when the model drifts, with automatic state pre-alignment to eliminate pose/velocity jumps at handover moments. Correction trajectories are timestamp-aligned with the original inference trajectory and saved in parallel, ready to feed back as training samples—every human correction becomes targeted teaching data.

Regarding existing tools: LeRobot is a training library and data-format standard, ROS is low-level communication and control abstraction, and VLASH / RTC are model- or algorithm-level point optimizations. All operate on the "model inference to action output" segment. Open-RAIL handles the work outside that segment—scheduling, smoothing, collection, evaluation—as a model-agnostic common layer.

What Runs on the Foundation

China Mobile's self-developed Mobile Star Kitchen robot runs on Open-RAIL. It has appeared at the Digital Summit, Shanghai MWC, and the China-ASEAN Expo, and runs in normal operation at a Hangzhou R&D campus with over 100 cups per day and a full ~90 seconds per latte from start to finished cup. Grabbing, placing, button-pressing, waiting, and cup transfer are all autonomous model decisions—no hand-written control logic for any step. It can also open a microwave door and set time and temperature.

Note: these figures come from the open-source project's own deployment descriptions, with no third-party capacity or stability audit. The 90-second pace is near commercial made-to-order coffee rhythm, which makes it a useful external reference point: model planning per se is not remarkable; what is remarkable is that every planned action lands stably.

Why the Open-Sourcer Matters

Embodied AI is an engineering discipline, and what blocks deployment is often not algorithms but infrastructure that produces no papers. China Mobile hosts Open-RAIL on its own Huanxin community, with GitHub and Gitee mirrors. That a state-owned enterprise is open-sourcing this layer is significant: the layer has strong public-good properties—hard for anyone to exclusively capture its value, hence easily skipped by all participants. Its positioning aligns with LeRobot and ROS: paving the ecosystem, not fencing territory. China Mobile's Hangzhou R&D center has also participated in authoring 5 national AI standards, which contextualizes its willingness to turn engineering conventions into public interfaces.

The Remaining Questions Are Operational, Not Technical

• Who will continuously contribute code after the release?
• How will the 10 model adaptations keep up with model generational turnover?
• Who will produce third-party data on long-term real-robot stability?
• What happens if the interface convention gets forked into incompatible versions?

A plain test for whether this path works: in six months, count how many companies submit merged pull requests, and how many of the 10 model adaptations were contributed by domestic vendors themselves.

References

1. The Paper, "Open-Sourced Globally! China Mobile's Open-RAIL" — https://www.thepaper.cn/newsDetail_forward_34080569 2. Sina Finance / IT Home, "Industry First: China Mobile Open-Sources Open-RAIL" — https://finance.sina.com.cn/tech/digi/2026-09-16/doc-iniryvce3073201.shtml 3. China.com, "China Mobile Open-Sources Open-RAIL to Bridge Embodied AI's Last Mile" — https://news.china.com/socialgd/10000169/20260916/49746085.html 4. Open-RAIL project page — https://cmcc-tao.github.io/open-rail/ 5. Open-RAIL repository (Gitee mirror) — https://gitee.com/cmcc-tao/open-RAIL