HERO: A self-improvement framework that bootstraps and consolidates robotic manipulation capabilities with zero human demonstrations.

Li et al. propose HERO, a hierarchical embodied agent that autonomously evolves manipulation capabilities without human demonstrations, using three modules—heuristic reasoning, exemplar reuse, and reflective execution—to gradually consolidate raw physical interaction experience into closed-loop visuomotor policies. The authors report substantial reductions in human intervention during robot data collection across various manipulation tasks while maintaining robust manipulation performance. The framework has risks such as unlabeled physical errors being distilled into the lower-level policy during the policy consolidation phase, and handoff failures at module transitions.