System-1 Planner: Direct, Fast & Adaptive Planning

• System-1 Planner is defined as a fast, reactive planning module that directly maps problem inputs to actions without explicit search.
• It employs direct plan generation algorithms and supervised learning to achieve low-latency responses ideal for real-time applications.
• Integrating with hybrid and multi-agent architectures, it balances speed with accuracy for domains such as robotics, language model planning, and materials discovery.

A System-1 Planner denotes a planning agent or module that generates plans or control actions in a fast, direct, and reactive manner, typically without explicit deliberation, global search, or backtracking. Rooted in dual-process theories of cognition, the System-1 label distinguishes such planners from deliberative "System-2" counterparts. In computational settings, System-1 Planners have been developed and deployed across diverse domains, ranging from real-time robotics to LLM–driven task decomposition. This entry surveys System-1 Planner architectures, algorithmic mechanisms, objective tradeoffs, representative applications, and properties such as flexibility, generalizability, and scalability.

1. High-Level Concepts and Core Characteristics

System-1 Planners are defined by their direct-generation nature: they produce a plan or immediate action sequence from an input problem description—be it a state pair $(s_0, s_g)$, a sensory observation, or a high-level task prompt—without performing stepwise internal search or explicit exploration of alternatives. This design allows for rapid, low-latency planning cycles suitable for real-time or high-throughput domains. The autonomy of System-1 Planners contrasts with System-2 Planners, which perform iterative search, branch-and-bound, or explicit state expansion.

Key attributes:

• Reactivity and Memorylessness: Plan or act based only on current input; typically lack reliance on global state or trajectory history (Lee et al., 2021).
• Direct Plan Generation: Map problem descriptions directly to action or plan outputs, as formalized by supervised learning objectives $$\mathcal{L}_{\mathrm{Sys1}} = -\sum_{(s_0,s_g),\mathcal{P}} \log p_\theta(\mathcal{P} \mid s_0,s_g)$$ (Saha et al., 2024).
• Minimal or No Search: Avoid explicit exploration, instead leveraging learned policy or heuristic knowledge.
• Speed and Computational Efficiency: Suitability for environments with tight timing constraints or large task backlogs, often at the expense of completeness or optimality.

2. System-1 Planners in LLM-Based Planning Frameworks

Recent planning systems leveraging LLMs have instantiated System-1 Planners as core components, especially in hybrid "System-1.x" and multi-agent architectures. The "System-1.x" framework (Saha et al., 2024) decomposes planning into three components: a Controller, a System-1 Planner (fast, direct), and a System-2 Planner (deliberative, search-based). The Controller decomposes problems into sub-goals, routing each to the most appropriate planner based on a hardness function and a user-specified hybridization factor $x$.

Within this system:

• System-1 Planner: Receives a sub-goal $(s_0,s_g)$ and emits a candidate plan using direct generation, with no intermediate search trace or backtracking.
• Training: Uses cross-entropy loss over ground-truth plans, with supervision given by valid solutions (optimal or suboptimal).
• Controllability: By tuning hybridization factor $x$, users can increase or decrease the proportion of sub-goals solved via System-1 (speed) versus System-2 (accuracy).

Experiments demonstrate that pure System-1 Planners offer fast plan generation with lower token cost (states explored), though with reduced accuracy on out-of-distribution or complex instances (Saha et al., 2024).

Sample Accuracy and Cost Results

Planner Type	Maze Accuracy	Maze #States	Blocksworld Accuracy	Blocksworld #States
System-1	48.7%	3.1	9.0%	4.5
System-2 (default)	93.7%	24.4	28.0%	55.5
System-1.x hybrid (with sub-goal, test-time)	96.7%	27.3	25.0%	38.3

System-1 Planners, while efficient, are typically less robust on difficult domains unless combined through systematic hybridization with search-based planners.

3. System-1 Planning in Hierarchical Multi-Agent Material Discovery

In the S1-MatAgent framework for materials discovery (Wang et al., 18 Sep 2025), the System-1 Planner serves as a high-level, LLM-based "brain" within a Planner–Executor architecture. Here, System-1 refers to the Planner component, which receives an inverse-design request, autoconfigures the problem decomposition, and dynamically orchestrates a team of Executor agents.

Operational structure:

• HTN Construction: The Planner recursively builds a hierarchical task network, breaking down root tasks via LLM-driven decomposition until all leaves are directly solvable by registered tools (Executors).
• Dynamic Tool Assignment: For each subtask, the Planner selects the minimal toolset required, instantiates a custom Executor, and passes along context and dependencies.
• Task Routing and Aggregation: Executors run subtasks in parallel where dependencies permit, passing results to the Planner, which monitors completion and assembles the final solution.

Pseudocode formalism for the decomposition-execution loop ensures minimal tool overhead, maximal concurrency, and completeness in subtask assignment.

4. Algorithmic and Mathematical Formulations

System-1 Planners generally optimize implicit objectives balancing directness, resource cost, and responsiveness. In S1-MatAgent (Wang et al., 18 Sep 2025), the Planner implicitly solves

$$\min_{\text{HTN},\,\pi} \biggl[\sum_{t\in \text{primitive}} C_{\text{tools}(t)} + \lambda\,T_{\text{makespan}}(\pi)\biggr]$$

subject to dependency and coverage constraints, where $C_{\text{tools}(t)}$ measures tool overhead, and $T_{\text{makespan}}(\pi)$ denotes execution makespan for schedule $\pi$. These costs are heuristically surfaced to the LLM (e.g., "prefer fewer tools", "parallelize steps").

In tasks involving optimization over combinatorial or continuous spaces, as in HEA catalyst composition, a gradient-based System-1 planning subroutine is embedded:

$$\nabla_{x}\,\mathcal{A}(x) = \frac{\partial f(x)}{\partial x}$$

with $$\mathcal{L}_{\mathrm{Sys1}} = -\sum_{(s_0,s_g),\mathcal{P}} \log p_\theta(\mathcal{P} \mid s_0,s_g)$$0 denoting compositional choices and $$\mathcal{L}_{\mathrm{Sys1}} = -\sum_{(s_0,s_g),\mathcal{P}} \log p_\theta(\mathcal{P} \mid s_0,s_g)$$1 the activity descriptor evaluated by a differentiable MLIP. Discrete composition modifications are guided by element-wise averaged gradients, repeatedly generating and evaluating candidate compositions for maximal activity improvement.

5. System-1 Planners in Reactive Robotic Control

A classical instantiation of System-1 planning appears in reactive, memoryless navigation for UAVs, exemplified by the RAPPIDS planner (Lee et al., 2021). At each camera frame, the planner samples candidate minimum-jerk (5th-order polynomial) trajectories seeded at the current state and ending within observable free space. The process proceeds as follows:

• No Map Retention: Only current depth frame informs planning; all out-of-view or occluded voxels are treated as obstacles.
• Free Space Representation: Candidate endpoints yield "free" rectangular pyramids, as defined by contiguous rectangles in depth image space projected into 3D.
• One-Step Utility Optimization: For each feasible primitive, utility $$\mathcal{L}_{\mathrm{Sys1}} = -\sum_{(s_0,s_g),\mathcal{P}} \log p_\theta(\mathcal{P} \mid s_0,s_g)$$2 is maximized, quantifying average speed toward the goal while rewarding proximity to endpoint.
• Fast Constraint Checks: Collision, velocity, and actuator constraints are checked; only the top-utility primitive is executed.

This loop achieves reactive, high-frequency (100 Hz) planning allowing UAV navigation in dense, cluttered environments without collision or loss of state estimation lock. Results indicate successful navigation through 30 m of dense forest at up to 2.7 m/s, with safety margins ≥0.3 m, and planning cycle times always below depth frame intervals.

6. Properties: Controllability, Flexibility, and Generalizability

System-1 Planners feature modularity and adaptability in their application and integration:

• Controllability: In hybrid architectures (Saha et al., 2024), the proportion of System-1 to System-2 execution is governed by an explicit hybridization factor $$\mathcal{L}_{\mathrm{Sys1}} = -\sum_{(s_0,s_g),\mathcal{P}} \log p_\theta(\mathcal{P} \mid s_0,s_g)$$3 at both training and test time, enabling performance–cost tradeoff through a user interface.
• Flexibility: Can be composed seamlessly with symbolic or neural System-2 planners, and extended to alternative decomposition heuristics or action representations without architectural rewrites.
• Generalizability: Demonstrated robustness to domain shifts and alternative search algorithms (A*, BFS, DFS), particularly when routed via a learned controller.
• Scalability: By design, System-1 Planners offload specialization to Executory agents or primitives, scale naturally to extensive tool registries or large system deployments, and minimize inter-module dependencies (Wang et al., 18 Sep 2025).

7. Representative Case Studies

• Material Discovery (S1-MatAgent): System-1 Planner orchestrates complete closed-loop HEA catalyst design: extracts and processes literature (1,231 formulas), analyzes composition, invokes LLM and code tools for recommendation, and leverages gradient-based MLIP optimization, yielding experimentally validated catalysts with up to 27.7% performance improvement (Wang et al., 18 Sep 2025).
• LLM Planning (System-1.x): System-1 Planner generates plans in Maze Navigation and Blocksworld domains, achieving competitive token efficiency and, when integrated with decomposition and System-2 search, state-of-the-art tradeoffs in accuracy and cost (Saha et al., 2024).
• UAV Obstacle Avoidance (RAPPIDS): Reactive System-1 Planner enables robust, low-compute trajectory selection based solely on the latest depth frame, ensuring collision-free flight in natural outdoor environments (Lee et al., 2021).

Summary Table: Key Features Across System-1 Planner Instantiations

Domain	Role of System-1 Planner	Key Benefit
Materials Science	LLM-driven HTN decomposition & scheduling	Full-cycle automation
LLM Planning	Direct plan generation	Low-token, fast inference
Robotics	Reactive trajectory generation	Real-time safety

The System-1 Planner paradigm encompasses a spectrum from highly reactive, memoryless action selection to high-level LLM-driven task allocation and decomposition, united by directness and the minimization of explicit deliberative search. This architecture supports rapid, scalable, and adaptable planning in domains characterized by uncertainty, size, or the need for real-time operation.