AMD ROCm pipeline enables CUDA-free VLA robot manipulation training

A new paper from Qing Yang and colleagues introduces Real2Sim2Real, an end-to-end AMD-accelerated pipeline for embodied AI manipulation. It's built entirely on the open ROCm software stack, eliminating the need for Nvidia’s CUDA. The team uses SmolVLA — a compact vision-language-action model — combined with the Genesis physics engine (for simulation) and 3D Gaussian Splatting (for reconstructing real scenes into synthetic data). The pipeline runs natively on Radeon AI PRO R9700 (RDNA4) and Radeon PRO W7900 (RDNA3.5) GPUs with PyTorch and is fully reproducible via the free Radeon Cloud Platform.

The authors demonstrate four progressive capabilities: (1) Sim-to-Real transfer where a VLA policy trained in simulation directly controls a physical Franka robotic arm; (2) a semantic 'one-of-three' object selection task driven by natural language; (3) Real2Sim synthetic data generation that fuses 3DGS reconstructions with the Genesis physics engine for training; and (4) large-scale reinforcement learning for quadruped and humanoid locomotion across multiple hardware platforms. The work aligns with industry leaders’ visions for Physical AI — Jensen Huang and Lisa Su have both called it the next big frontier. This pipeline proves that a fully open, AMD-only ecosystem can rival CUDA-based workflows for robot learning.