A Skill-based Hierarchical Learning Framework for Long-Horizon Robotic Manipulation Tasks

In long-horizon tasks, robots often struggle with vast exploration spaces. To address this, we propose SkillRo, a hierarchical skill-based learning framework. We first use a conditional generative model to pre-train a generalizable skill repository from heterogeneous datasets, enabling efficient skill inference across tasks. This reduces the need for task-specific training. Then, we integrate a small number of task-specific demonstrations to refine the skill selection process, narrowing the exploration space. This allows the robot to quickly acquire task-solving strategies. We evaluate SkillRo on two challenging long-horizon tasks: a standard task and a distribution mismatch task. Experimental results show that SkillRo significantly improves learning efficiency, task complexity handling, and generalization to unseen tasks compared to existing baselines.