GRAIL (Generating Humanoid Loco-Manipulation from 3D Assets and Video Priors) is an NVIDIA research pipeline that fully digitally synthesizes training data for humanoid robots. It composes 3D assets, simulator-ready scenes, and robot-proportioned characters, then combines them with video foundation model priors (including Kling and an optional local MiniMax-H3 backend) to generate metric 4D human-object interaction trajectories. Those trajectories are retargeted onto a Unitree G1 humanoid and used to train task-general policies for pick-up, whole-body manipulation, sitting, and terrain traversal such as stairs and slopes. The project reports that policies trained purely on GRAIL-generated data transfer to real-world object pick-up and stair-climbing using egocentric vision, without collecting any real robot data. Code, documentation, task-general tracking policy checkpoints, and a companion Hugging Face dataset (PhysicalAI-Robotics-Locomanipulation-GRAIL) have been released, along with Docker images for reproducing the pipeline. The codebase covers the full loop: 3D asset generation, 2D human-object-interaction generation, 4D reconstruction, retargeting, tracking, and data export/visualization tools.