AI Agents Learn to Plan Subgoals by Watching the Whole Task Unfold

Autonomous agents that follow natural-language instructions often struggle once a task stretches beyond a few steps. A hierarchical policy framework called EOPBiSE (Episodic On-the-Fly Planner and Bifurcate Subgoal Executor), described in an open-access paper published in Complex & Intelligent Systems, addresses that weakness by grounding its planning in what the agent actually sees as the episode unfolds. The core problem the researchers identified is a discrepancy between plan and status.

When a plan is generated only from the language instruction, it cannot adapt if the environment deviates from what the plan assumed, so the agent may keep pursuing subgoals that no longer match the real-time state of the world. EOPBiSE tackles this with a planner that draws on both the linguistic instruction and holistic visual observations taken from the task's entire episode history, allowing subgoal predictions to be updated in an adaptive, on-the-fly manner rather than fixed in advance. Execution is handled by a separate component, the bifurcate subgoal executor.

By splitting the subgoal information it receives, the executor is designed to minimize the gap between what was planned and what is actually happening during execution, translating abstract intermediate goals into concrete actions more reliably. The team evaluated the framework on ALFRED, a benchmark of complex household tasks carried out from an egocentric visual perspective. According to the paper, EOPBiSE effectively handled a wide range of long-horizon tasks and showed generalization capabilities in unseen scenarios, meaning it maintained performance when tested on environment configurations it had not been trained on.

The reported results come from a single benchmark of simulated household tasks, and the abstract-level claims concern task handling and generalization within that setting. The planner's use of episodic visual history is the key departure from prior approaches, tying each predicted subgoal to accumulated evidence about the task's progress instead of to the instruction text alone, with the reduction of plan-status discrepancy described as the mechanism behind the framework's performance.

Subject of Research: A hierarchical AI policy that plans and executes subgoals for long-horizon instruction following tasks

Article Title: Learning to plan and execute subgoals with a hierarchical policy for long-horizon instruction following tasks

Article References: Zhuang, B., Zhang, C., Zhang, M., & Hu, Z. (2026). Learning to plan and execute subgoals with a hierarchical policy for long-horizon instruction following tasks. Complex & Intelligent Systems. https://doi.org/10.1007/s40747-026-02416-y

Image Credits: AI Generated

DOI: 10.1007/s40747-026-02416-y

Keywords: EOPBiSE, hierarchical policy, autonomous agents, instruction following, subgoal planning, ALFRED benchmark, long-horizon tasks, egocentric vision, machine learning, AI planning, Learning, plan