Heterogeneous Recursive Planning

• Heterogeneous recursive planning is a framework integrating recursive task decomposition with diverse agent actions and representations.
• It employs hierarchical tree structures that mix primitive and generalized actions using both model-based search and policy-guided expansion.
• The methodology improves multi-agent mission performance through distributed multi-objective optimization and adaptive scheduling.

Heterogeneous recursive planning is a class of methodologies in AI and robotics characterized by recursive, hierarchical decomposition of tasks or missions, where the resulting subtasks are heterogeneous in their execution requirements, representations, or agent capabilities. This paradigm fuses flexible hierarchical search with mechanisms for exploiting policy learning, heterogeneity in agents or actions, and multi-stage optimization cycles. Core domains of application include multi-robot mission planning, autonomous exploration, and hierarchical control for long-horizon sequential decision problems.

1. Fundamentals and Formal Definitions

Heterogeneous recursive planning integrates hierarchical task decomposition with heterogeneity at multiple levels, such as agent capabilities, action sets, or elemental planning operators. Tasks are recursively decomposed into subtasks or actions that can differ in structure—e.g., composite vs. atomic—while also supporting assignments to distinct agents based on capabilities or environmental constraints (Ferreira et al., 2021).

Formally, such frameworks operate via a recursive procedure (e.g., plan(a, s)), where "a" can be primitive actions or generalized actions (GAs). At each recursion, a new planning subproblem is instantiated with its own distinct search space and constraints, reflecting underlying heterogeneity. Central to this recursion is the ability to mix primitives and learned subroutines (GAs), incorporating multiple action types, reward structures, and planning modalities (Redlich, 2024).

In multi-agent scenarios, such as HEHA for multi-robot exploration (Yang et al., 5 Oct 2025), heterogeneity manifests as varying robot traversal capabilities, assignment constraints, and cost profiles for edges in planning graphs. The broader architecture supports hierarchical planners, evolutionary schedulers, and hybridized search-policy agents.

2. Hierarchical Task Decomposition and Tree Representations

A typical implementation models the mission or task as a hierarchical tree, where:

• Task nodes (composite): Non-leaf nodes representing tasks requiring further decomposition.
• Action nodes (atomic): Leaf nodes denoting primitive or indivisible execution steps.

Operators such as AND (all children required) and XOR (exactly one child required) drive the decomposition logic. Each child may be annotated, for each agent $i$, with a triple $(q_a(i), d_a(i), c_a(i))$, denoting quality, duration, and cost, respectively. Precedence constraints (edges) can enforce scheduling dependencies such as $finish(a) < start(b)$ (Ferreira et al., 2021).

The recursive task decomposition process employs algorithms (e.g., generate_alternatives) that, for each composite task $t$, recursively solicit alternatives for each subtask and aggregate them according to the respective logical operator:

Function generate_alternatives(task t): if t is an action a: return [[a]] all_alt = [] for each child sub in subtasks(t): sub_alt = generate_alternatives(sub) all_alt.append(sub_alt) if Q(t)==AND: t_alt = cartesian_product(all_alt) elif Q(t)==XOR: t_alt = [] for set in all_alt: for alt in set: t_alt.append([alt]) if |t_alt| > μ: scores = [sc(alt) for alt in t_alt] pick top μ by score return t_alt

Score functions combine weighted sums or Pareto criteria over action quality, duration, and cost, enabling multi-objective optimization.

3. Heterogeneity in Agents, Actions, and Policies

Heterogeneity is reflected through assignment constraints, differentiated agent action sets, and context-sensitive scheduling or planning operators. In multi-robot planning, clusters are assigned based on robot-specific capabilities (e.g., air, legged, wheeled), encoding both individual traversal constraints and global mission distribution (Yang et al., 5 Oct 2025).

Recursive tree planners (RTP) extend heterogeneity to the algorithmic level by allowing arbitrary mixing of primitives and generalized actions throughout all levels of the planning hierarchy (Redlich, 2024). A GA is defined as a structure containing an internal action set $A_g$ (which may itself be a mix of primitives and lower-level GAs), a stochastic policy $p_g(a|s)$, subgoals $S_g$, and planning parameters $\gamma_g$.

Distinct agent priorities or region exclusivity are managed through mechanisms such as priority assignment, which deterministically allocates agents to tasks based on class precedence, and cost adjustment formulas tailored for region-exclusive frontiers.

4. Recursive Search, Policy Integration, and Mixed Modes

The search phase in heterogeneous recursive planning encompasses both model-based enumeration and policy-based selection. In RTP, regimes span from pure-planner (policy scores ignored) to pure-policy rollout (no search) and hybrid “near-greedy” expansion, wherein stochastic policies prioritize search locally but model-based search explores alternative paths (Redlich, 2024).

The generalized Dijkstra search is adapted to handle heterogeneous action sets and policy-guided expansion with rewiring for optimality and optional heuristic coupling for A*-like behavior. Hierarchical calls (plan(a,g,s)) trigger recursive sub-planning, naturally handling the expansion and refinement of both primitive and complex actions.

Policies are learned via supervised imitation from successful plans, producing task-specific $p_g(a|s)$. These can exploit invariances such as background or object-number invariance, facilitating policy generalization and zero-shot transfer across distinct problem instances.

5. Multi-Objective Optimization and Scheduling

Scheduling and execution in heterogeneous recursive planning are often realized through distributed metaheuristics, multi-objective genetic algorithms, and Pareto fronts (Ferreira et al., 2021). Chromosomes encode action allocations as robot-specific routes, with intra- and inter-schedule precedence enforced. Fitness is defined via makespan (e.g., $C_{max}$), cost, and quality, with nondomination ranks and density estimates used to drive multiplicity in the Pareto set.

In HEHA, global planning proceeds via the PEAF algorithm—a label-setting, partial-expansion A* variant—with joint state-space search, dominance pruning, and focal window suboptimality guarantees: any solution $\pi$ satisfies $max_i C(\pi^i) \leq (1+\epsilon) max_i C^*(\pi^i)$, where $\pi^*$ is optimal (Yang et al., 5 Oct 2025).

Local planners further refine each agent’s route, respecting heterogeneity by adjusting costs for region-exclusive exploration and enforcing priority-based assignment. Execution is iteratively staged, allowing responsive reaction to newly discovered subproblems and dynamic adaptation as the mission progresses.

6. Empirical Evaluation and Application Domains

Empirical studies demonstrate marked performance improvements in both synthetic and real-world scenarios:

• In multi-robot mission planning, recursive tree search combined with distributed evolutionary scheduling adapts flexibly to robot availability and multi-objective optimization criteria, as shown in an automated greenhouse use case (Ferreira et al., 2021).
• The HEHA architecture for unknown environment exploration achieves up to 30.2% reduction in exploration time and 31.8% shorter total path length compared to frontier-based NBVP baselines, with bounded suboptimality via $\epsilon$-controlled focal search (Yang et al., 5 Oct 2025).
• RTP enables rapid policy bootstrapping, zero-shot transfer, and combinatorial speedup in Box2D and MuJoCo domains, facilitating orders-of-magnitude acceleration over pure-planning baselines and robust generalization to new object counts and dynamic conditions (Redlich, 2024).

In all cases, recursive refinement, feedback-driven decomposition, and the exploitation of domain-specific heterogeneity result in scalable performance for long-horizon, complex tasks.

7. Significance, Generalization, and Open Directions

Heterogeneous recursive planning unifies recursive task decomposition, hybrid agent/action assignment, dynamic scheduling, and policy-guided planning, establishing a general paradigm applicable across agents with disparate capabilities and hierarchical task structures. Its domain independence, support for pure-policy and pure-planning extremes, and facilitation of virtuous cycles of planning, imitation, and policy improvement enable its use in adaptive writing agents (Xiong et al., 11 Mar 2025), multi-robot exploration, and hierarchical reinforcement learning.

This suggests the flexibility and generality of heterogeneous recursive planning architectures can be leveraged for both expert systems and learning-enabled control in domains requiring high adaptability, policy transfer, and compositionality. A plausible implication is the convergence of model-based and data-driven planning strategies through recursive hybridization, allowing future systems to robustly address high-dimensional, multimodal environments.