AIO Planner: Integrated AI Scheduling

• AIO Planner is an integrated AI planning system combining LLMs, symbolic schedulers, and iterative refinement loops to generate robust, multi-domain plans.
• It handles both vague natural language constraints and explicit requirements using LLM-driven implicit reasoning alongside symbolic extraction for strict feasibility.
• The system is applied across scheduling, educational design, and autonomous driving, and is evaluated using metrics like correctness, latency, and user satisfaction.

An AIO (All-In-One) Planner denotes a class of integrated AI-driven planning systems capable of managing complex, multi-step, and often multi-domain planning objectives in domains such as daily scheduling, software agents, interdisciplinary educational design, and autonomous driving. AIO planners architecturally combine symbolic components, LLMs, constraint handling, iterative refinement loops, and multi-modal data sources to produce robust, user-centric plans that handle both explicit and vague, natural-language constraints.

1. System Architectures and Core Design Patterns

The heterogeneity of AIO Planner design reflects their deployment environment and planning targets. Common foundational elements include:

(a) LLM-driven Plan Generation:

LLM-based planners (e.g., LLMPlan) formalize the input as tuples $e = (d, L)$, where $d$ is a description and $L = \{l_1,\ldots,l_m\}$ are constraints. The model is prompted with natural-language task descriptions, loosely constrained event lists (including vague requirements), and one-shot schedule exemplars. Plan output is a set of scheduled events $c_i = (e_i, \mathsf{start}_i, \mathsf{end}_i)$ (Li et al., 2023).

(b) Modular Symbolic Schedulers:

Complementary to LLM backbones, symbolic planners (e.g., SymPlan) translate explicit constraints into formal clauses (e.g., “event: duration < 30 minutes”) that instantiate objects in a Simple Temporal Network (STN). Symbolic solvers utilize shortest-path and backtracking for feasibility and constraint satisfaction (Li et al., 2023).

(c) Two-stage Planner-Action Agents:

On-device frameworks like Octo-planner decouple plan formation and action execution via a “Planner agent” (e.g., Phi-3 Mini) producing ordered sub-steps, and an “Action agent” (Octopus model) executing API calls per plan task. This separation allows for resource-efficient, modular deployment on edge devices (Chen et al., 2024).

(d) Iterative, Multi-Component Prompting Pipelines:

AI-driven education platforms instantiate a multi-stage planning flow—such as IDPplanner’s ten-component scaffold, guiding users through sequential, template-driven plan development, each stage driven by separate LLM prompts and iteratively refined via human-in-the-loop edits (Liow et al., 17 Oct 2025).

(e) End-to-end Multi-modal Integration:

In autonomous motion planning, architectures like Int2Planner jointly encode spatio-temporal context, route structure, and multi-agent state, providing a unified prediction-planning loop sensitive to real-world interaction uncertainty (Chen et al., 22 Jan 2025).

2. Natural-Language and Symbolic Constraint Handling

AIO Planners demonstrate contrasting strategies for the ingestion and operationalization of planning constraints:

• LLM-driven Implicit Constraint Reasoning:

In LLMPlan, all user-specified constraints—including those expressed in vague or incomplete natural language—are fed verbatim into the prompt without explicit parsing. The LLM leverages its pretrained knowledge to interpret and embed these requirements within the generated plan, yielding strong support for ambiguous queries (e.g., “I’d like frequent breaks in the afternoon”) (Li et al., 2023).

• Explicit Symbolic Extraction and Translation:

Planners like SymPlan first pass natural-language messages through lightweight extractors, producing canonical “before/after” or “duration” clauses that are then mapped to STN edges. This approach enables hard constraint enforcement and exhaustive feasibility checks, at the expense of flexibility in handling ambiguous or underspecified user intents (Li et al., 2023).

• Domain-specific Prompt Engineering:

AIO educational planners feature prompt builders encoding syllabus parameters, taxonomies, and class profiles as variables in meticulously designed prompt templates—allowing precise curriculum alignment while remaining framework-agnostic (Liow et al., 17 Oct 2025).

3. Self-Reflection and Iterative Refinement Mechanisms

Modern AIO Planners incorporate self-critique or validation modules that operate in an iterative loop to systematically eliminate errors and violations:

(a) LLM Self-Reflection Loop:

After initial plan generation, auxiliary checkers are invoked: (i) constraint-checkers for hard violation detection (e.g., event overlap), (ii) missing-event detectors, and (iii) overlap detectors. Violations are verbalized into feedback and appended to the next prompt. The LLM runs iteratively, with performance improvements typically saturating after $K=3{-}5$ cycles (Li et al., 2023).

(b) Hybrid Human–AI Feedback:

In educational settings, each plan component is subject to teacher review and critique. The feedback is formalized into a prompt, with guided re-generation proceeding until qualitative criteria are satisfied (Liow et al., 17 Oct 2025).

4. Evaluation Metrics and Quantitative Performance

AIO Planner evaluation targets correctness, commonsense alignment, and end-user satisfaction. Standardized metrics include:

• Correctness Metrics:
  • Event Coverage (CO): Percentage of input events scheduled
  • Non-Overlap (NO): Percentage of non-overlapping events
  • Satisfaction rates for Duration Constraints (DC), Order Constraints (OC), Start/End Constraints (SEC)
• Commonsense Metrics:
  • Violations of curated commonsense bounds (e.g., “breakfast < 60 min”)
  • Deviations from time-of-day norms
• Efficiency and Resource Use:
  • Planning accuracy, latency, memory, and power draw on representative hardware (see Octo-planner benchmarks in the table below)

Summary of System-Level Benchmarks (Chen et al., 2024)

Method	Accuracy	Latency (ms)	RAM (MB)	Power (W)
Phi-3 Mini (full FT)	98.1%	400	2,048	2.1
LoRA (r64, α256)	85.1%	380	1,500	1.9
Multi-LoRA (2 domains)	82.2%	390	1,600	2.0

User-interactive evaluations consistently reveal that LLM-centric planners outpace symbolic baselines in handling vague constraints and user refinements, approximately doubling observed user satisfaction (e.g., LLMPlan 70.5% vs. SymPlan 40.4% positive feedback) (Li et al., 2023). Commonsense violations are reduced by LLM planners across both synthetic and real-user studies.

5. Domain Applications and Generalization Strategies

AIO Planners have been successfully instantiated in diverse domains:

• Personal Assistant and Scheduling:

Day planning systems (LLMPlan, SymPlan) for adaptive, constraint-aware schedule generation, responsive to free-form user input (Li et al., 2023).

• On-Device Multi-domain Agents:

Octo-planner demonstrates the feasibility of accurate planning on resource-constrained devices, merging multiple functional domains via LoRA adapter strategies and enabling plug-and-play extensibility (Chen et al., 2024).

• Interdisciplinary Educational Design:

IDPplanner operationalizes a ten-component flow for interdisciplinary project generation, allowing deep curriculum alignment and customizable scaffolding. The system’s prompt library and feedback loop enhance structure, sequencing, and idea diversity (Liow et al., 17 Oct 2025).

• Integrated Motion Planning in Robotics:

Int2Planner leverages route intention points to tightly couple interactive prediction and planning under uncertainty, yielding closed-loop performance gains in urban autonomous driving (Chen et al., 22 Jan 2025).

Framework-agnostic design is a recurring principle; modular templates, taxonomy mappings, and internal plug-in infrastructure (syllabus, rubrics, learner models) allow domain adaptation and localized deployment (Liow et al., 17 Oct 2025). Techniques such as multi-LoRA adapter merging facilitate efficient horizontal scaling across new APIs and function sets (Chen et al., 2024).

6. Hybrid and Human-in-the-Loop Planning Paradigms

Empirical results emphasize the superiority of hybrid workflows in complex or context-sensitive tasks. Human users oversee contextualization (e.g., local examples, accommodations), inject refinements, and select optimal plan variants. System design recommendations for AIO Planners across domains include:

• Integrated override capabilities and editable templates at each planning step
• Transparent presentation of generation prompts for reproducible outputs
• Iteration prompts and critique stages to deepen integration or expand plan detail
• Explicit audit trails and prompt history recording to support policy compliance and traceability (Liow et al., 17 Oct 2025)

7. Limitations, Open Challenges, and Future Directions

Complex plan manipulations—such as higher-order transformations (“split meeting into two slots”)—remain failure modes for both LLM-based and symbolic systems, with success rates near 50% (Li et al., 2023). Scalability to high-dimensional action spaces and fully cross-domain reasoning require further innovation, possibly via deeper chain-of-thought prompting, specialized editing modules, or incorporation of real-time feedback data.

AIO Planners represent a convergence of LLM advances, symbolic scheduling, and interactive refinement paradigms, achieving near-optimal constraint satisfaction while offering significant gains in commonsense planning, user satisfaction, modularity, and platform efficiency.