Principles2Plan: LLM-Guided System for Operationalising Ethical Principles into Plans

AI summary

Overview

Research area: Computational Machine Ethics (CME) applied to automated planning, at the intersection of large language models (LLMs), human-computer interaction, and classical AI planning.

Technical level: Intermediate. The paper is a system demonstration, so the concepts are presented accessibly, but it assumes familiarity with automated planning, PDDL, and the notion of action costs. The authors themselves state that intended users are assumed to have a basic understanding of automated planning and familiarity with PDDL and PDDL-Ethical.

Scope: The paper presents Principles2Plan, an interactive research prototype in which a human domain expert and an LLM collaborate to turn high-level ethical principles into context-specific, operationalisable rules that guide a classical planner toward ethically-informed plans.

What This Paper Is About

Robots acting around people need to respect ethical principles such as beneficence and privacy, but those principles are abstract and their correct application depends heavily on context. Existing automated planning tools offer little support for this: manually writing ethical rules is labour-intensive and highly context-specific, and fully automated ethical planning is difficult because interpreting abstract principles is nuanced and often controversial. The goal of this work is an interactive system where a user supplies the planning domain, problem details, and high-level principles, and an LLM generates operationalisable ethical rules that the user can review, refine, prioritise, and hand to a planner to produce ethically-informed plans.

Key Contributions

1. A prototype system (Principles2Plan) that integrates an interactive, user-facing interface with the human-in-the-loop ethical planning pipeline introduced by Zhong et al. (2026), letting users generate, refine, and prioritise context-specific ethical rules.
2. A four-step guided workflow spanning dedicated pages: providing input, reviewing and prioritising generated rules, reviewing generated PDDL-Ethical code, and comparing the resulting plans.
3. Human-LLM collaboration for principle operationalisation: the system makes explicit how abstract principles are turned into actionable rules, with the LLM supplying explanations of its reasoning and the human retaining authority to add, remove, modify, and weight rules.
4. A claim of novelty: to the authors' knowledge, no prior system supports users in generating principle-grounded rules for classical planning contexts, or collaborative human-LLM refinement and operationalisation of ethical principles.

Main Findings

• No prior comparable system: The authors state that while prior work lies at the intersection of users, LLMs, and automated planning, no existing system supports collaborative human-LLM refinement and operationalisation of ethical principles. Favier et al. (2025) provides an interface but targets general natural language constraints in PDDL3; Zhong et al. (2026) targets ethics but lacks a user-facing interface. Principles2Plan addresses this gap.
• Ethical principles are context-dependent: The paper illustrates that the same principle (beneficence) can justify different actions depending on context — an unauthorised shortcut may be justified for a passenger needing urgent medical attention, whereas for a leisure trip following standard traffic rules may be preferable to avoid unnecessary risk.
• Reported performance figures (from prior work, not this prototype): The method was evaluated with DeepSeek-R1-Distill-Llama-70B in Zhong et al. (2025), achieving Sentence-BERT similarity of 0.82 for generated rules and a code generation success rate of 82.2%. The authors describe these results as "not exceptional" but indicative of a promising direction.
• First-of-its-kind prototype may need high human intervention: Because this is the first implemented prototype of its kind, the authors acknowledge it may require more human intervention in its current form.
• Usability barrier acknowledged: Intended users are unlikely to have technical knowledge of planning and PDDL, and minimising the need for such expertise remains a challenge for future work.
• Design choices: Rules include ethical features representing positive or negative ethical characteristics (e.g., dishonesty as a negative feature), which users prioritise by assigning a significance level of 1–5.

Methodology in Plain English

The system is a walkthrough of four pages, each corresponding to a stage of the workflow:

1. Input page. The user uploads problem.pddl and domain.pddl files, specifies the initial state, states assumptions about the problem or domain, enters the high-level ethical principles to guide rule generation, and selects a preferred model. The system then prompts the LLM to produce context-specific ethical rules in natural language, following a structure defined in Zhong et al. (2026), each with associated ethical features. To support experimentation, the system ships with example problems across three ethically-sensitive domains: autonomous vehicles, elderly care, and firefighting/rescue.
2. Ethical rules editor. Because LLM-generated rules may be inconsistent or imperfect, the user reviews them, adds missing rules, removes inappropriate ones, and modifies existing ones, aided by LLM explanations of the reasoning behind each rule. The user then assigns a significance level (1–5) to each ethical feature, with the system highlighting positive and negative features for easy adjustment.
3. Code editor. The prioritised rules are fed back to the LLM to produce syntax-highlighted PDDL-Ethical code. The user reviews it, and the code is transpiled into raw PDDL with action costs using the method from Jedwabny (2022). A side view of the natural language rules supports cross-checking for correctness and consistency. The result is submitted to a domain-independent classical planner, Fast Downward.
4. Output plan page. The plan produced with ethical rules and a second plan produced by the same planner using the original problem and domain files are displayed side-by-side, so users can directly evaluate the impact of the ethical rules.

The design deliberately keeps the human in the loop at every stage — the LLM proposes, the human disposes — in contrast to fully automated or fully manual rule encoding.

Why This Matters

Impact on research: The paper positions itself in Computational Machine Ethics, which the authors divide into top-down approaches (rules specified in advance, transparent but not adaptable), bottom-up approaches (data-driven, flexible but less interpretable), and hybrid approaches (which combine strengths but typically still need extensive manual encoding effort). By using LLMs to reduce that manual effort within a classical planning context, Principles2Plan offers a practical research direction and a hands-on platform for experimenting with interactive ethical decision-making.

Real-world applications:

• Autonomous vehicles — deciding when urgences, such as a passenger needing urgent medical attention, might justify deviations from standard traffic rules, and when they do not.
• Elderly care robotics — balancing beneficence and privacy when assisting people in care settings.
• Firefighting and rescue — planning under ethically-sensitive constraints where goals and principles may conflict.
• General service and assistive robots operating in human environments, where explicit, reviewable ethical rules are preferable to opaque behaviour.

Industry relevance: The system produces ethically-informed plans from standard PDDL inputs and an off-the-shelf classical planner (Fast Downward), rather than requiring bespoke ethical reasoning engines. For organisations deploying robots in regulated or safety-critical settings, the side-by-side comparison of ethical and non-ethical plans offers a transparent audit-style view of how principles change behaviour — a property that purely bottom-up or data-driven approaches do not easily provide.

Future Directions

1. Improving the human-LLM collaboration balance — the authors state future work will enhance collaboration through iterative dialogue and suggestions, with the aim of improving the balance between human effort and LLM output.
2. Reducing the PDDL expertise barrier — the paper explicitly identifies minimising the need for users to have technical knowledge of planning and PDDL as a challenge for future work.
3. Closing the performance gap — the reported metrics (Sentence-BERT similarity of 0.82 and code generation success rate of 82.2%), which the authors call "not exceptional," suggest scope for better rule generation and code generation.
4. Reducing required human intervention — the authors acknowledge the prototype may need more human intervention in its current form and express optimism that future iterations will improve on this.

Target Audience

Domain experts working in ethically-sensitive domains, AI ethics and robotics researchers, and anyone interested in the intersection of ethics, LLMs, and automated planning. The authors also note that because the process includes reviewing code and generating plans, users are assumed to have a basic understanding of automated planning and familiarity with PDDL and PDDL-Ethical, though they recognise intended users are unlikely to have such technical knowledge — a gap the work aims to eventually narrow.

Authors’ abstract

Ethical awareness is critical for robots operating in human environments, yet existing automated planning tools provide little support. Manually specifying ethical rules is labour-intensive and highly context-specific. We present Principles2Plan, an interactive research prototype demonstrating how a human and a Large Language Model (LLM) can collaborate to produce context-sensitive ethical rules and guide automated planning. A domain expert provides the planning domain, problem details, and relevant high-level principles such as beneficence and privacy. The system generates operationalisable ethical rules consistent with these principles, which the user can review, prioritise, and supply to a planner to produce ethically-informed plans. To our knowledge, no prior system supports users in generating principle-grounded rules for classical planning contexts. Principles2Plan showcases the potential of human-LLM collaboration for making ethical automated planning more practical and feasible.