SARM dual-head reward model

Stage-and-progress reward model trained with LeRobot SARM (annotation_mode=dual) on bimanual SO-101 demonstrations. Overhead camera only (observation.images.left_top). Final checkpoint is step 2500, batch 128.