T3 Planner: A Self-Correcting LLM Framework for Robotic Motion Planning with Temporal Logic

TL;DR

T3 Planner addresses the challenge of translating natural language instructions into executable robotic motion under complex spatio-temporal constraints by introducing a self-correcting, three-stage LLM-driven planning framework augmented with formal STL verification. The cascaded Task, Time, and Trajectory planners operate in a closed loop, with logical, temporal, and motional verifications ensuring feasibility and robustness, and a knowledge-distilled, lightweight Qwen3-4B variant enables efficient deployment. Experimental results across Household, Chip, and Navigation scenarios show significant performance gains over a translation-based baseline (AutoTAMP), with high success rates (often above 90%) even for compact models. The approach demonstrates strong practical impact by reducing infeasible plans, enabling safer, more reliable execution in real-world robotics, and offering a scalable path to deployment via distillation.

Abstract

Translating natural language instructions into executable motion plans is a fundamental challenge in robotics. Traditional approaches are typically constrained by their reliance on domain-specific expertise to customize planners, and often struggle with spatio-temporal couplings that usually lead to infeasible motions or discrepancies between task planning and motion execution. Despite the proficiency of Large Language Models (LLMs) in high-level semantic reasoning, hallucination could result in infeasible motion plans. In this paper, we introduce the T3 Planner, an LLM-enabled robotic motion planning framework that self-corrects it output with formal methods. The framework decomposes spatio-temporal task constraints via three cascaded modules, each of which stimulates an LLM to generate candidate trajectory sequences and examines their feasibility via a Signal Temporal Logic (STL) verifier until one that satisfies complex spatial, temporal, and logical constraints is found.Experiments across different scenarios show that T3 Planner significantly outperforms the baselines. The required reasoning can be distilled into a lightweight Qwen3-4B model that enables efficient deployment. All supplementary materials are accessible at https://github.com/leeejia/T3_Planner.

Figures (7)

• Figure 1: Illustration of $T^3$ Planner
• Figure 2: Prompt templetes of $T^3$ Planner
• Figure 3: $T^3$ Planner produces feasible motion trajectories. (a) In the household scenario, the instruction is "Visit each room once, except for the Bedroom and Living Room, and complete the whole process within 30 seconds." (b) In the chip scenario, the instruction is "Strive to reach all the goals. At the same time, always avoid touching the walls. Complete the process in less than 30 seconds." (c) In the navigation scenario, the instruction is "Seek to reach every single goal. Also, always make sure not to touch the obstacles. Complete the process in less than 30 seconds."
• Figure 4: Computational efficiency of task planner
• Figure 5: Computational efficiency of time planner