Hierarchical Planning

Train agents to decompose complex tasks into hierarchical subgoals, plan at multiple levels of abstraction, and execute plans through lower-level policies — enabling long-horizon task completion. Hierarchical planning trains agents to operate at multiple levels of abstraction. A high-level planner decomposes a task into subgoals (a plan), and lower-level policies execute each subgoal. The high-level planner receives the overall task and the current world state, and outputs a sequence of subgoals. Each subgoal is passed to a lower-level policy trained to achieve that specific subgoal. Training alternates between high-level plan optimization and low-level skill refinement, often using a combination of supervised learning on demonstrations and RL on task completion.