Hierarchical Planning

• Hierarchical planning is a decision-making framework that decomposes high-level objectives into manageable subgoals to facilitate long-horizon task execution.
• It employs methodologies such as HTN planning and neural generator–refiner architectures to bridge abstract goals with concrete, executable actions.
• Applications in robotics, resource allocation, and high-dimensional planning demonstrate its effectiveness in enhancing scalability, robustness, and computational efficiency.

Hierarchical planning is a set of methodologies in artificial intelligence, robotics, and cognitive science that address the challenge of decision-making across multiple abstraction levels. It leverages domain structure by decomposing high-level goals into more tractable subgoals or subtasks, each of which can be further refined until concrete, executable actions are specified. This approach supports long-horizon reasoning, sample and computational efficiency, and robust generalization by reusing sub-solutions and enabling modular representations. Hierarchical planning underpins both symbolic systems (such as Hierarchical Task Network (HTN) planning) and contemporary machine learning frameworks that integrate foundation models for language, vision, and control.

1. Formal Definitions and Models of Hierarchical Planning

At its core, hierarchical planning operates over a space of goals or tasks $G$, a set of states $S$, and a set of primitive actions $A$. The agent's objective is to construct a hierarchical plan, typically a sequence or tree of abstract and primitive steps, that transforms an initial state $s_0$ into one that satisfies a specified goal.

Formal probabilistic model: In recent neural hierarchical planning systems, the full plan likelihood for solving a language-specified goal $g$ from an initial visual observation $x_{1,1}$ is factorized as

$$p_\Theta(W, \{\tau_x^i\}, \{\tau_a^i\} \mid g, x_{1,1}) = \prod_{i=1}^N p_\theta(w_i|g) \cdot p_\phi(\tau_x^i|w_i, x_{i,1}) \cdot \prod_{t=1}^{T-1} p_\psi(a_{i,t}|x_{i,t}, x_{i,t+1})$$

where $W = (w_1, \ldots, w_N)$ are symbolic subgoals, $\tau_x^i$ are image (video) trajectories associated with each subgoal, and $\tau_a^i$ are corresponding action sequences (Ajay et al., 2023).

Recursive decomposition: Symbolic and programmatic frameworks define decompositions or methods that map an abstract task into a sequence of sub-tasks, down to primitive actions. For example, in HTN planning,

• A task network $tn$ is a partially ordered set of tasks (primitive or compound).
• A method specifies how to refine a compound task using a subnetwork.
• A solution is a sequence of primitive actions obtained by repeated decomposition, respecting ordering and constraints (Georgievski et al., 2014, Höller et al., 2019).

In learning-driven systems, library-based methods store decompositions indexed by task and may employ linguistic hints or learnable proposal functions to guide decomposition selection (Cano et al., 2023).

2. Algorithmic Principles and Architectures

Hierarchical planning algorithms typically follow one or more of the following paradigms:

Task network decomposition: Symbolic planners apply recursive application of methods to expand compound tasks, examining ordering, preconditions, and domain constraints. This yields partially or totally ordered plans, depending on the model (Georgievski et al., 2014, Höller et al., 2019).

Generator–Refiner architectures: Modern neural planners employ separate modules for high-level subgoal generation and for subgoal refinement, often jointly trained or refined iteratively. For example, in HiP (Ajay et al., 2023):

• A LLM proposes symbolic subgoals.
• Each subgoal is grounded by a video diffusion model synthesizing visual plans from current context.
• Actions to execute video plans are inferred by an inverse dynamics model. Consistency is enforced via iterative feedback between modules.

Skill abstraction: Agent capabilities are structured as a set of skills, each associated with an effect and precondition; planning proceeds by composing and recursively invoking these. Markov Intent Process-based approaches further formalize skill policies that can recursively delegate subplans (Lai et al., 2020).

Diffusion-based hierarchical planning: High-level "jumpy" planners subsample the state trajectory, proposing sparse subgoals, while low-level diffusion models fill in detailed transitions between subgoals. The result is substantial reductions in planning and computation required for long-horizon tasks (Chen et al., 2024, Pertsch et al., 2020).

Curriculum-driven skill acquisition: Some frameworks explicitly learn or refine high-level skills based on a goal curriculum, promoting incrementally harder problems and automatic abstraction creation (Morere et al., 2019, Cano et al., 2023).

3. Applications, Empirical Results, and Benchmarks

Hierarchical planning has been applied to a range of domains, with empirical evaluation focusing on efficiency, scalability, generalization, and solution quality:

• Robotics and manipulation: HiP achieves 72–85% success on long-horizon tabletop manipulation tasks without oracle subgoals; baselines such as single-level transformers or action diffusion models attain 20–75% (Ajay et al., 2023).
• Resource allocation: City-scale emergency response planning is decomposed into region-level allocation and intra-region dispatch, leading to reductions in mean response time up to 21.6 seconds compared to static baselines in real-world EMS data from Nashville (Pettet et al., 2020, Pettet et al., 2021).
• Planning in high-dimensional spaces: Annotated skeleton-guided planners, e.g., HASP, construct sparse roadmaps and rapidly identify near-optimal paths in cluttered, multi-query robot environments. HASP achieves path lengths within 5–10% of state-of-the-art while performing an order of magnitude fewer collision checks (Uwacu et al., 2022).
• Generalized, modular, or recursive policy synthesis: Hierarchical finite-state controllers (h-FSCs) can compactly represent families of recursive and parameterized plans, with exponential savings in controller size (Segovia-Aguas et al., 2019, Stefansson et al., 15 Jan 2025).

4. Theoretical Foundations and Complexity

Hierarchical planning can dramatically reduce search complexity by factorizing the planning problem across abstraction levels. Key theoretical results and design choices include:

• Complexity reduction: Hierarchical decomposition reduces the effective branching factor and search depth from exponential in problem size to exponential in the number of abstraction layers, enabling tractable optimization in otherwise intractable state spaces (e.g., up to $2^{100}$ states) (Morere et al., 2019).
• Expressivity: HTN formalisms are strictly more expressive than STRIPS-style planners, supporting partially ordered tasks, user-controlled decomposition, and domain constraints. However, plan existence is undecidable for partially ordered HTNs in the general case (Georgievski et al., 2014).
• Soundness and completeness: Analysis of summary information propagation in hierarchical plan libraries shows that preconditions for abstract steps are complete (i.e., every successful refinement requires the precondition to hold), while effect summaries are sound for non-recursive hierarchies (Silva et al., 2017).
• Learning-theoretic efficiency: Hierarchical models achieve faster learning of reusable skills and better adaptation to changes in dynamics or task objectives (Cano et al., 2023, Consul et al., 2021).

5. Human and Cognitive Perspectives

Hierarchical planning aligns with empirical and theoretical models of human problem-solving:

• Human program induction: Experimental evidence shows that humans favor plans that simultaneously minimize action cost, program length (minimum description length, MDL), and, crucially, promote subroutine reuse beyond mere compressibility, consistent with nonparametric Bayesian grammar induction models (Correa et al., 2023).
• Cognitively informed strategy discovery: Hierarchical decomposition enables the design of planning strategies for humans that outperform both non-hierarchical approaches and previously known heuristics; teaching such strategies results in significant human performance improvements on complex, multistep decision tasks (Consul et al., 2021).

6. Extensions, Limitations, and Future Directions

Contemporary and emerging research highlights several avenues and constraints:

• Automated subgoal and library induction: Ongoing advances seek to automate the discovery and adaptation of subgoal libraries, skill hierarchies, and landmarks, mitigating manual domain engineering (Cano et al., 2023, Li et al., 26 Aug 2025).
• Benchmarking and languages: The advent of HDDL standardizes hierarchical domain modeling and enables fair comparison across hierarchical planning systems, facilitating the development of large-scale, mixed-paradigm benchmarks (Höller et al., 2019).
• Integration with foundation models: Cutting-edge architectures combine symbolic and generative models for hierarchical planning, enforcing cross-modality consistency (e.g., between language, vision, and control) via iterative refinement and feedback (Ajay et al., 2023).
• Adaptive planning in dynamic settings: Hierarchical planning in city-scale CPS, robotics, and tracking tasks demonstrates robust adaptation through periodic high-level reallocation and local online replanning, with incremental updates in the presence of environmental changes (Pettet et al., 2020, Yuan et al., 12 Oct 2025, Stefansson et al., 15 Jan 2025).
• Limitations: Performance may depend on the granularity of abstraction, the completeness and robustness of libraries or skeletons, parameter tuning, and computational bottlenecks in training large generative models. Research continues on extending frameworks to open-ended goal sets, richer programming constructs, and real-time, on-device deployment (Cano et al., 2023, Chen et al., 2024, Uwacu et al., 2022).

7. Connections and Taxonomy

Hierarchical planning interconnects multiple research traditions:

Model/Framework	Key Abstraction	Solution Structure	Typical Application Domains
HTN Planning	Task networks	Partially-ordered plans	Automated planning, workflow composition
Hierarchical Skill/Option Discovery	Skill library	Flat or nested policies	RL, robotics, navigation
Neural Generator-Refiner Models	Latent subgoals	Abstract→concrete mapping	Long-horizon vision-based manipulation
Library-based Program Synthesis	Program library	Recursively instantiated	Program induction, tool composition
h-FSCs / HFSMs	Sub-controller	Modular or recursive FSM	Large-scale system control, generalized planning

This taxonomy illustrates the diversity of approaches and their convergence around hierarchical decomposition as a mechanism for scalability, modularity, and generalization (Georgievski et al., 2014, Höller et al., 2019, Segovia-Aguas et al., 2019, Stefansson et al., 15 Jan 2025).