AI
New Hybrid Approach Combines AI Language Models with Logic Programming for Robot Task Planning
Researchers have developed a system that merges the broad knowledge of large language models with structured logic programming to make robotic task plans actually executable.
Researchers have introduced a novel approach called CLMASP that bridges two complementary AI paradigms to improve how robots understand and execute complex tasks. The system pairs the world knowledge and planning intuitions of large language models with the formal reasoning capabilities of Answer Set Programming, a specialized logic system designed to capture robot action constraints.
The Problem: Plans That Don't Work#
Large language models excel at drawing on vast amounts of training data and can sketch out reasonable high-level plans for unfamiliar tasks. Yet they struggle to ensure those plans are actually executable given a robot's real-world limitations—its available actions, physical constraints, and timing requirements. Without precise grounding in what a specific machine can actually do, LLM-generated instructions often fail when a robot tries to carry them out.
How CLMASP Works#
The system operates in three stages. First, an LLM generates a rough plan outline based on the task description—a skeleton that captures the intent and major steps. Next, this skeleton is adapted to the particular scenario using a vector database, enriching it with relevant contextual details. Finally, the plan passes through an Answer Set Programming module that encodes the robot's specific action knowledge and constraints, transforming the abstract LLM output into a concrete, executable sequence.
This layered design lets the LLM contribute creative problem-solving while ASP ensures all proposed actions respect the robot's actual capabilities.
Dramatic Improvement in Execution#
In tests on the VirtualHome simulation platform, CLMASP achieved a dramatic leap in real-world applicability. Standalone LLM approaches produced executable plans less than 2% of the time. CLMASP boosted that figure to over 90%, demonstrating that combining symbolic reasoning with neural language understanding can yield far more robust task planning.
