Cortical Module (High-Level Goal Planner)

• Cortical module is a modular computational component for abstract decision making in AI, analogous to the prefrontal cortex in biological systems.
• It decomposes long-term goals into subgoals using hierarchical planning methods ranging from symbolic approaches to LLM-based strategies.
• Its modular design enhances task generalization and efficiency by decoupling strategic planning from reactive execution, though subgoal validation remains challenging.

A cortical module, in the context of artificial high-level goal planning, is a modular computational component that assumes the role of long-horizon, abstract decision making—analogous to the prefrontal cortex in biological systems. In embodied and multi-modal agents, “cortical module” specifically refers to the system responsible for decomposing goals, selecting intermediate targets, and orchestrating the sequence and structure of sub-tasks presented to downstream low-level controllers. The defining properties of the cortical module are explicit separation from reactive or motor modules, periodic invocation on long planning horizons, and hierarchical or map-based representations supporting abstraction/generalization across tasks and domains.

1. Modular Role, Scope, and Interface

The cortical module operates as a distinct high-level planner that explicitly decouples long-term goal decomposition and route prediction from short-term action selection and reactive behaviors.

• Inputs: Typically receives task-specific goal information (e.g., a panoramic goal image, PDDL-encoded target state, or natural language instruction), recent sensory perceptions (multi-view RGB, depth, occupancy maps), and often a spatial or semantic map encoding both environment topology and prior agent trajectories (Wu et al., 2021, Kwon et al., 2024, Ge et al., 12 Jan 2025, Song et al., 2022).
• Outputs: Produces a discrete or continuous subgoal—in geometric (ℝ² region), symbolic (ordered subgoal set), or skill/instructional (text plan) form—presented directly to a subordinate planner or skill/execution module, which further decomposes it to primitive (motor-control or API) actions (Wu et al., 2021, Chen et al., 23 Apr 2025, S et al., 2024).
• Update Policy: Invoked at set intervals (after every K low-level steps) and when new perceptual or task information necessitates strategic plan revision (Wu et al., 2021, Chen et al., 23 Apr 2025, Song et al., 2022).

This architectural decoupling enables targeted training and improved generalization: the cortical module learns and outputs plan structure, leaving detailed physical interaction, motion, or execution to optimized downstream blocks. Inter-module interfaces are clearly defined—often through explicit message-passing or modular RL conventions (e.g., plan-vector or subgoal-tuple broadcast) (Si et al., 7 Oct 2025, Dalal, 10 Mar 2025).

2. Core Algorithmic Mechanisms

Three dominant algorithmic patterns arise in modern cortical module design:

1. Goal Decomposition and Subgoal Generation:
   • Classic planners encode the problem in PDDL or MDP form, decompose the main goal S* into ordered, tractable subgoals {S*_1, ..., S*_n} via symbolic partitioning, LLM world modeling, or region-based abstraction (Kwon et al., 2024, Ge et al., 12 Jan 2025, Rens, 3 Jan 2025). LLM approaches use few-shot prompting to output textual subgoal sequences (Song et al., 2022, Chen et al., 23 Apr 2025, Dalal, 10 Mar 2025).
   • Hierarchical RL variants construct semantic or spatial subgoal graphs and optimize over sequences of goal-conditioned policies (Li et al., 2022, Nasiriany et al., 2019).
2. Search, Planning, and Policy Selection:
   • Search-based modules use explicit tree/planner constructs (e.g., MCTS or BFS over product automata) to rank possible subgoal trajectories based on learned or model-based value estimates (Rens, 3 Jan 2025, Ge et al., 12 Jan 2025, Li et al., 2022).
   • In value-based schemes, a Q-network or a learned critic scores subgoal candidates, often using deep CNNs or VAEs over high-dimensional state and latent spaces (Núñez-Molina et al., 2020, Nasiriany et al., 2019).
3. Cognitive and Modular Orchestration:
   • Cortical modules can instantiate sub-functions matching cognitive neuroscience models: task decomposition (prefrontal), action selection (dlPFC), constraint/check monitors (ACC), internal model-based prediction or state evaluation (OFC), and gating orchestration (aPFC) (Webb et al., 2023). Such modular architectures wire LLM or neural planners into decentralized, recurrent loops, allowing distributed reasoning and robust plan validation.

3. Representative Architectures and Key Designs

A selection of implemented cortical modules in the literature illustrates both algorithmic variety and unifying principles:

Paper	Goal Planner Input	Internal Representation / Model	Output/Interface	Planning Method
(Wu et al., 2021)	{I_g, oₜ, mₜ}	ResNet18 × 2 + Conv map encoder	gₜ ∈ ℝ² (map target loc.)	PPO-optimized RL policy
(Kwon et al., 2024)	{PDDL, goal + nl}	LLM subgoal generator, size estimator	Ordered subgoal list, selects symbolic/MCTS	LLM+symbolic+MCTS hybrid
(Ge et al., 12 Jan 2025)	{nl prompt}	LTL formula → DFA → product TS	Region sequence R	BFS/search over abstract graph
(Si et al., 7 Oct 2025)	{task instruction}	Plan text, evaluated by executors	High-level plan (text)	Plan synth. + PPO/GRPO RL stage
(Rens, 3 Jan 2025)	{goal set, state}	Hierarchical plan tree; GCPs via RL	HLA sequence (goal-conditioned policies)	MCTS with learned GCPs
(Song et al., 2022)	{instr., objects, history}	Plan as subgoal tuples	τ=[g₁,…,g_T]: ordered symbolic subgoals	Few-shot LLM + grounded updates

All architectures define explicit inputs, modular representations for planning/subgoal selection, and consistent handoffs (subgoal, policy, or plan) to reactive/low-level planners.

4. Learning, Training, and Optimization Paradigms

Cortical modules are trained using diverse methods tailored to their abstraction level and required generalization:

• Supervised pre-training and few-shot prompting: Cold-start via high-quality plan sets from advanced LLMs, filtered by executor performance (Si et al., 7 Oct 2025, Song et al., 2022). Fine-tuned with cross-entropy over plan tokens and optionally chain-of-thought reasoning.
• Reinforcement learning (policy optimization): RL-based planners use PPO or group-relative PPO, with custom objective functions reflecting task-completion improvement or efficiency relative to plan-free execution (Wu et al., 2021, Si et al., 7 Oct 2025). Alternately, modular RL setups may train goal-conditioned policies and value networks jointly with variational state abstractions (Nasiriany et al., 2019, Li et al., 2022).
• Model-based and hybrid planning: Symbolic planners (Fast-Downward, classical LTL search) are invoked by LLM submodules, with meta-control over decomposition and plan assignment based on expected complexity or value (Kwon et al., 2024, Ge et al., 12 Jan 2025).
• Value/policy function approximators: CNN-based DQL for subgoal selection, variational autoencoders (for feasible state/goal latent spaces), and critic networks for evaluating predicted subgoal achievement (Núñez-Molina et al., 2020, Li et al., 2022, Nasiriany et al., 2019).

Empirically, data efficiency and generalization to hard tasks are paramount; LLM- and modular-planner-based cortical modules consistently outperform end-to-end RL on multi-step, branching, or compositional tasks (Song et al., 2022, Chen et al., 23 Apr 2025, Dalal, 10 Mar 2025).

5. Cognitive Analogies and Neuroscientific Parallels

Multiple research works explicitly link the cortical module to biological executive control (Chen et al., 23 Apr 2025, Webb et al., 2023), describing:

• Functional mapping: High-level goal decomposition and plan maintenance mapped to the anterior PFC, subgoal selection to central PFC, plan evaluation/constraint monitoring to the ACC, outcome/state prediction to the OFC, and modular orchestration to the PFC and subcortical loops.
• Modular, hierarchical separation: The explicit division between strategy (goal planner) and action execution (reactive/motor modules) mirrors functional separation between prefrontal and motor cortices; cognitive loops of attend, perceive, store, plan, and execute are realized by separate programmatic modules (S et al., 2024).
• Abstraction and plan sequencing: Long-range, temporally extended plan representation is enabled by compact abstract state spaces (e.g., VAE latent codes for feasible observations (Nasiriany et al., 2019)), direct analogs of neural population codes and hierarchical chunking in human PFC (Webb et al., 2023, Rens, 3 Jan 2025).

While explicit neural plausibility (e.g., learning via local synaptic updates or spike-driven credit assignment) is not fully realized, several works discuss replacing non-local RL updates (PPO, cross-entropy) with more biologically plausible actor–critic or Hebbian schemes (Si et al., 7 Oct 2025, Núñez-Molina et al., 2020).

6. Empirical Performance and Limitations

• Benchmark performance: Modular cortical planners yield substantial improvements in success rate (≥88–100% on standard planning/task domains (Kwon et al., 2024, Chen et al., 23 Apr 2025, Dalal, 10 Mar 2025)), planning efficiency (10–100× speedup over flat planners in hard domains (Kwon et al., 2024, Núñez-Molina et al., 2020)), and closed-loop accuracy in navigation and manipulation (Wu et al., 2021, Wang et al., 2023).
• Ablation studies: Removal of the cortical module or flattening of hierarchy sharply degrades multi-step task success and increases planning time (Kwon et al., 2024, Chen et al., 23 Apr 2025).
• Limitations: Current systems require accurate map/building modules and are sensitive to subgoal definition quality; LLM-driven planners remain prone to hallucinations or over-commitment to ungrounded subgoals if not connected to perceptual or physical world state (Dalal, 10 Mar 2025, Song et al., 2022). Trade-offs exist between planning optimality, computational overhead, and real-time adaptability (e.g., open-loop LLM plans versus feedback-driven controllers) (Dalal, 10 Mar 2025).

7. Extensions, Open Problems, and Future Directions

• Scaling and generalization: Unified frameworks that combine large-scale policy learning with plug-and-play modular planners demonstrate robust transfer and sample efficiency (Dalal, 10 Mar 2025, Song et al., 2022).
• Hybrid symbolic–neural planners: Neuro-symbolic cortical modules that allocate subproblems to either declarative symbolic planners or neural LLM/MCTS hybrids optimize both speed and accuracy (Kwon et al., 2024).
• Cognitive plausibility and continual learning: Augmentation of current planners with attention mechanisms, recurrent memory, and local credit signals may further close the gap to biological systems (Núñez-Molina et al., 2020, Si et al., 7 Oct 2025, Webb et al., 2023).
• Grounding and feedback: Improved closed-loop plan repair, subgoal validation, and error-driven re-planning under partial observability remain unsolved for LLM-based and learned cortical modules (Song et al., 2022, Dalal, 10 Mar 2025).

Cortical modules—characterized by their explicit, modular, and hierarchical goal planning functionality—have thus become a unifying principle across state-of-the-art embodied control, agentic language systems, and modular RL architectures. Their precise formulation, signalling interfaces, and cognitive parallels are central topics in contemporary AI planning research.