LEGS: Fine-Tuning Teleop-Free VLAs for Humanoid Loco-manipulation in an Embodied Gaussian Splatting World
LEGS is a hybrid simulator that creates labeled humanoid locomotion data without needing human teleoperation or video. It combines a photorealistic 3D Gaussian Splatting background with a mesh foreground representing robots and objects. This synthetic data allows Vision-Language-Action (VLA) policies to be fine-tuned on simulators, achieving performance equal to or better than those trained using real human demonstrations.