CLMASP: Coupling Large Language Models with Answer Set Programming for Robotic Task Planning

The paper presents CLMASP, a method that integrates Large Language Models (LLMs) with Answer Set Programming (ASP) to significantly improve the executability of robotic task plans in complex environments, achieving over 90% success rate on the VirtualHome platform by refining LLM-generated plan skeletons with ASP to ensure they are practical and adhere to specific constraints.

Abstract

Large Language Models (LLMs) possess extensive foundational knowledge and moderate reasoning abilities, making them suitable for general task planning in open-world scenarios. However, it is challenging to ground a LLM-generated plan to be executable for the specified robot with certain restrictions. This paper introduces CLMASP, an approach that couples LLMs with Answer Set Programming (ASP) to overcome the limitations, where ASP is a non-monotonic logic programming formalism renowned for its capacity to represent and reason about a robot's action knowledge. CLMASP initiates with a LLM generating a basic skeleton plan, which is subsequently tailored to the specific scenario using a vector database. This plan is then refined by an ASP program with a robot's action knowledge, which integrates implementation details into the skeleton, grounding the LLM's abstract outputs in practical robot contexts. Our experiments conducted on the VirtualHome platform demonstrate CLMASP's efficacy. Compared to the baseline executable rate of under 2% with LLM approaches, CLMASP significantly improves this to over 90%.

Method

The authors used a methodology called CLMASP (Coupling Large Language Models with Answer Set Programming for Robotic Task Planning), which involves a two-tiered approach to robotic task planning. In the first tier, a Large Language Model (LLM) generates a basic plan skeleton based on natural language instructions. This skeleton is then refined through syntactic self-refinement and semantic referring-grounding to correct errors and ensure semantic accuracy. In the second tier, Answer Set Programming (ASP) is used to further refine the plan by formalizing the robot's action model into causal rules and translating them into ASP rules, which are then used to generate an executable plan for the robot. This methodology allows for the efficient generation of executable plans without the need for extensive training or fine-tuning, and it significantly improves the executability of robot plans, as demonstrated by the high success rate achieved in experiments on the VirtualHome platform.

Main Finding

The authors discovered that by integrating Large Language Models (LLMs) with Answer Set Programming (ASP), they could significantly improve the executability of robotic task plans in complex environments. Their CLMASP methodology involved using LLMs to generate a basic plan skeleton based on natural language instructions, which was then refined through syntactic self-refinement and semantic referring-grounding. This skeleton was further refined by ASP to ensure it was executable by the robot. The authors found that their method achieved an executable rate of over 90% on the VirtualHome platform, a substantial improvement over the under 2% success rate of baseline LLM approaches. This demonstrated the effectiveness of their approach in grounding abstract plans in practical robot actions and adhering to specific constraints that are often overlooked by LLMs alone.

Conclusion

The conclusion of the paper is that the CLMASP methodology, which integrates Large Language Models (LLMs) with Answer Set Programming (ASP), significantly improves the executability of robotic task plans in complex environments. The authors demonstrate this through their experiments on the VirtualHome platform, where CLMASP achieved an executable rate of over 90%, a substantial improvement over the baseline LLM approaches which had an executable rate of under 2%. The paper concludes that CLMASP's two-tiered approach, which involves generating a basic plan skeleton with LLMs and then refining it with ASP, is an effective and efficient method for robotic task planning that does not require extensive training or fine-tuning. The authors suggest that their findings have implications for the development of cognitive user interfaces for device control and that their approach could be generalized to other complex scenarios.

Keywords

Large Language Models, Answer Set Programming, Robotic Task Planning, VirtualHome platform, Executability, Task Planning, Knowledge Representation, Logical Reasoning, Natural Language Instructions, Action Model, Causal Rules, ASP Solver, Skeleton Plan, Syntactic Self-Refinement, Semantic Referring-Grounding, Vector Database, Exogenous Knowledge, Action Knowledge, Planning Problem, Goal Achievement Rate, CLMASP Methodology, Cognitive User Interface, Device Control, Open-World Scenarios, Non-Monotonic Logic Programming, Formalism, Action Knowledge, Implementation Details, Practical Robot Contexts, Efficacy, Executable Rate, Task Decomposition, Action Model, State Transition, Causal Model, ASP Rules, Planning Problem, Skeleton Plan, Trajectory, Causal Theory, Fluent-Specification, Subtask, Action Model, Planning Problem, Initial State, Goal State, Action Description, Logic Program, Answer Set Semantics, Planning Performance, Executability, Goal Achievement Rate, Environmental Flexibility, Closed-Source LLM Constraints, Automating ASP Program Generation, Ontology-Based Knowledge Graph, Inductive Logic Techniques.