Introduction
Robotic task planning in complex, real-world environments like kitchens requires a delicate balance between common-sense reasoning and strict logical consistency. Traditional logic-based approaches, such as Answer Set Programming (ASP), excel at enforcing physical constraints—for instance, ensuring a robot does not attempt to wash a shirt unless it is first placed in the machine. However, ASP faces a significant "grounding bottleneck": as the number of objects in an environment grows, the computational complexity of translating abstract rules into specific instances explodes, often making real-time planning impossible.
Conversely, Large Language Models (LLMs) possess vast common-sense knowledge but frequently struggle with technical executability. They may hallucinate objects or omit critical intermediate steps. This paper introduces CLMASP, a framework that synergizes the intuitive "rough solutions" of LLMs with the rigorous verification of ASP to achieve efficient and reliable robotic planning.
The CLMASP Framework
The CLMASP framework addresses the grounding bottleneck through three integrated strategies:
- Skeleton Plan Generation: Instead of an exhaustive search, the LLM generates a high-level "skeleton" (e.g., "Find shirt -> Wash shirt"). The ASP solver then uses this as a heuristic guide, focusing only on the low-level actions required to fulfill this sequence.
- Loop-Admissible Reductions: This is the core technical innovation. The authors mathematically prove that the planning domain can be pruned. By identifying which objects are irrelevant to the current task (e.g., ignoring a lawnmower when the task is washing dishes), the system reduces the number of constants the ASP solver must process.
- Semantic Grounding Refinement (SGR): To bridge the gap between natural language and robot databases, SGR uses text embeddings to map LLM-generated terms (like "kitchen_table") to specific environment IDs (like "table_01").
Experimental Evaluation
The framework was evaluated using the VirtualHome (VH) simulator, featuring environments with 250–300 objects and 3,000 spatial relationships. The researchers tested 269 distinct tasks using GPT-3.5, GPT-4o, and Llama 3.1-8B.
Key Results
The integration of LLM heuristics and formal domain reduction yielded dramatic performance improvements across several key metrics:
- Computational Efficiency: While standard ASP planning often exceeded a 2-hour timeout, CLMASP reduced average planning times to under 5 seconds. This is largely due to the reduction in grounding complexity.
- Improved Solvability and Executability: By providing a guided search space, the solvability rate ($R_{sol}$) increased by up to 86%. Furthermore, the executability of the resulting plans reached nearly 90%, far surpassing the performance of LLMs acting alone, which often had success rates below 10%.
Conclusion and Future Work
CLMASP demonstrates that LLMs are most effective when used as heuristic guides for formal solvers rather than as standalone planners. By mathematically reducing the search space based on LLM intuition, the grounding bottleneck is effectively bypassed without sacrificing logical soundness. Future research will explore dynamic replanning in non-static environments and improving the capacity of LLMs to handle even larger, more complex environmental maps directly.
Summary updated