Open-RAIL: China Mobile's Real-Time Execution Layer Between VLA Models and Humanoids
China Mobile open-sourced Open-RAIL, formerly VLA-RAIL, a real-time execution layer that sits between Physical AI models and robot bodies. The problem it targets is that a VLA or world action model thinks in action chunks while the robot needs fast continuous motion, so Open-RAIL runs a three-thread asynchronous pipeline (observation / inference / control) with two-level intra- and inter-chunk online smoothing that eliminates a 30-50x frequency gap and cuts joint acceleration std from 10+ to 0.1 rad/s2, with smoothing done at the framework layer so diffusion, flow matching and autoregressive models all work unchanged. A server-client split with non-overlapping dependency trees lets embedded robots drive large models and switch between device, edge and cloud with zero code changes. A lightweight RobotBase hardware abstraction layer plus unified action_layout indexing has adapted 4 heterogeneous robots (AgiBot G1, China Mobile Lingxi Ti5 T170C, NAVIAI-WA2 and a LeRobot mock simulation backend), and 10 VLA/WAM models across 7 families (ACT, GR00T N1/N1.5/N1.6, RDT-1B, SmolVLA, AgiBot GO-1, pi0/pi0.5, TAO) integrate in under 100 lines. Inference, data collection and evaluation are built into every run, writing LeRobot-style Parquet, and three modes cover pure inference, pure teleop and hybrid with pause and human intervention. Apache-2.0.
