SafeGen-LLM: Enhancing Safety Generalization in Task Planning for Robotic Systems

SafeGen-LLM is a post-training framework designed to enhance safety generalization in robotic task planning by addressing the limitations of classical planners, reinforcement learning (RL)-based methods, and base large language models (LLMs) . Classical planners, such as Fast Downward and Metric-FF, suffer from poor scalability and rigid input/output formats, while RL-based approaches exhibit limited generalization and require extensive environment interactions . Base LLMs, although capable of handling diverse inputs like natural language and PDDL specifications, lack safety guarantees and often produce plans that violate safety constraints .

To overcome these challenges, SafeGen-LLM introduces a two-stage training framework: Supervised Fine-Tuning (SFT) on a constraint-compliant planning dataset to learn planning syntax and semantics, followed by Group Relative Policy Optimization (GRPO) guided by fine-grained reward machines derived from formal verification . The reward machines enforce safety alignment, while curriculum learning helps manage complex tasks . A multi-domain Planning Domain Definition Language 3 (PDDL3) benchmark with explicit safety constraints is used to train and evaluate the model, enabling systematic assessment of safety compliance and generalization .

Experiments show that SafeGen-LLM achieves strong safety generalization across domains and input formats, outperforming frontier proprietary models in safety-aware planning despite having fewer parameters . When integrated with assurance frameworks like SafePilot, SafeGen-LLM can further guarantee safety while improving planning efficiency . This approach enables the generation of task plans that not only satisfy existing safety requirements but also generalize effectively to novel safety properties in previously unseen domains .

SafeGen-LLM: Enhancing Safety Generalization in Task Planning for Robotic Systems

SafeGen-LLM addresses the critical challenge of generating task plans for robotic systems that are both adaptable to novel environments and strictly compliant with safety constraints. Traditional robotic planning methods often struggle to balance these competing demands: classical symbolic planners are logically rigorous but brittle when facing unseen object properties or domain shifts, while reinforcement learning (RL) agents can adapt but frequently fail to generalize safety guarantees outside of their training distributions. SafeGen-LLM leverages the semantic reasoning capabilities of Large Language Models (LLMs) to bridge this gap, constructing a framework that synthesizes high-level task plans capable of operating across diverse domains. The system specifically focuses on "safe generalization," ensuring that a robot can effectively interact with objects possessing novel properties—such as new tools or obstacles—without violating predefined safety protocols.

The key contribution of this work is the integration of LLM-driven reasoning with a safety verification layer that filters and refines generated plans. Unlike standard prompting techniques that may hallucinate unsafe actions, SafeGen-LLM employs mechanisms to ground the model's output in safety logic, allowing it to infer safe manipulation strategies based on object descriptions rather than relying on exhaustive training data. The research demonstrates that this approach significantly outperforms both classical planning algorithms and state-of-the-art RL methods in scenarios requiring generalization. Empirical results show that SafeGen-LLM maintains a high success rate in task completion while drastically reducing safety violations in out-of-distribution (OOD) test cases.

This research is significant because it directly targets a primary bottleneck in the deployment of autonomous robots in unstructured, human-centric environments. For robots to be reliably deployed in settings like homes, hospitals, or dynamic industrial sites, they must handle unexpected variables without compromising safety. SafeGen-LLM provides a viable pathway toward "safe autonomy," proving that the generative power of LLMs can be harnessed to enforce rigorous safety constraints. By demonstrating that semantic understanding can be translated into safe physical actions, this work lays the groundwork for more robust, trustworthy, and versatile robotic systems capable of operating in the real world.

Summary of SafeGen-LLM: Enhancing Safety Generalization in Task Planning for Robotic Systems

The paper introduces SafeGen-LLM, a novel framework that leverages large language models (LLMs) to generate safe, generalizable task plans for robotic systems. Unlike traditional rule-based or reinforcement learning (RL) approaches, SafeGen-LLM exploits the zero-shot generalization capabilities of LLMs to adapt to novel environments and task properties without explicit retraining. The method combines safety-constrained planning with LLM-based reasoning, enabling robots to generate plans that comply with safety constraints while handling unseen scenarios effectively.

A key contribution of the work is its empirical validation against classical planning (e.g., PDDL-based solvers) and RL-based methods (e.g., CQL, PPO), demonstrating superior generalization and safety compliance in tasks with novel property distributions. The paper highlights how SafeGen-LLM mitigates the brittleness of classical planners and the sample inefficiency of RL methods by framing task planning as a natural language-guided search problem. This approach is particularly relevant for real-world deployment, where robots must operate in dynamic, unpredictable settings while adhering to safety constraints.

Why it matters: This work advances the state-of-the-art in safe and adaptive robotic planning, bridging the gap between AI-driven reasoning and real-world robustness. By integrating LLMs into the planning pipeline, SafeGen-LLM offers a scalable solution for zero-shot adaptation in safety-critical domains, from warehouse automation to assistive robotics. The findings underscore the potential of LLM-augmented control to enhance reliability in robotic systems without extensive domain-specific engineering.

Source: [arXiv:2402.24235](https://arxiv.org/abs/2402.24235)