Multi-Phase Multi-Objective Dexterous Manipulation with Adaptive Hierarchical Curriculum

Dexterous manipulation tasks usually have multiple objectives. The priorities of these objectives may vary at different phases a task. Current methods do not consider the objective priority and its change during task, making robot hard time or even fail to learn good policy. In this work, we develop novel Adaptive Hierarchical Curriculum guide with prioritized Our method determines learning process updates sequence adapt changing phases. A smooth transition function is developed mitigate effects on stability when updating sequence. proposed validated in multi-objective task JACO arm which needs manipulate target obstacles surrounded. simulation physical experiment results show that outperforms baseline 92.5% success rate 40 tests average takes 36.4% less finish