FluxVLA Engine and Alibaba Cloud PAI Team Up to Accelerate Embodied Intelligence into the Physical World

LimX Dynamics, an AI‑driven humanoid robot company, announced a deep collaboration with Alibaba Cloud’s Platform of Artificial Intelligence (PAI), migrating its training workloads to the PAI platform.

By moving to PAI, LimX achieved flexible GPU quota management, intelligent job scheduling, high GPU utilization, self‑healing failures, fine‑grained data permission control and hot‑cold data flow. In internal tests the migration yielded a 10 % increase in overall training efficiency and a 17 % reduction in operational complexity, establishing a robust AI infrastructure for rapid model iteration.

PAI is a one‑stop AI engineering platform that provides end‑to‑end capabilities—from data processing to model training and inference. Its proprietary distributed training acceleration engine supports thousands of GPUs and offers data, compute, optimizer and scheduling optimizations for large‑scale models such as large language models, world models, robot reinforcement learning, and autonomous driving.

LimX also released its self‑developed, standardized VLA model engineering base, the FluxVLA Engine, as an open‑source image on PAI. The engine follows a “unified configuration, standard interfaces, modular decoupling, accelerated deployment” design, covering the full chain from data handling, training, simulation to real‑robot deployment.

FluxVLA Engine supports mainstream VLA models such as OpenVLA, GR00T, and PI0/PI0.5, and is compatible with hardware like TRON 2 and Aloha. On an RTX 5090 GPU the GR00T‑N1.5 model reaches 42.8 Hz inference speed, significantly lowering the engineering barrier for bringing embodied‑intelligence models from the lab to physical deployment.

The engine is offered as a standardized image in the PAI‑DSW module, allowing users to launch the full environment with a single click, fine‑tune, and deploy their own embodied‑intelligence models.

This collaboration aims to break traditional robotics development barriers, enabling small and medium‑size teams to share leading research outcomes without rebuilding AI infrastructure, and to deepen exploration of world‑model research so that embodied‑intelligence technologies can move out of the laboratory and integrate into everyday life.