Cost-Aware LLM-Based Task and Motion Planning with Learned Feasibility Checks

Abstract

Large Language Models (LLMs) have emerged as powerful zero-shot task planners for robotic manipulation, reducing the need for manually specified symbolic models in classical Task and Motion Planning (TAMP). However, in closed-loop LLM–planner systems, motion planning (MP) is often the primary source of feasibility feedback. This results in excessive MP queries and LLM API calls, becoming a major efficiency bottleneck.We address this gap by introducing a lightweight, modular pre-check stage in front of the motion planner. Before an LLM-proposed action (e.g., Pick/Place) is passed to MP, it is validated by (i) a fast analytical axis-aligned bounding boxes (AABB)-based geometric check and (ii) an optional learned feasibility predictor trained from offline MP rollouts. The resulting structured feedback (e.g., collision reports and feasibility scores) is fed back to the LLM, encouraging more cost-aware decisions between cheap parameter adjustment and expensive backtracking. We evaluate methods using both success rate and a weighted planning cost that aggregates MP calls and LLM calls.On the LLM3 box-packing benchmark, our pre-check framework reduces planning overhead (MP calls and LLM calls) while maintaining comparable success rates to the baseline, and yields the largest efficiency gains in harder settings with tighter packing constraints.