🚀 Adaptive Iterative Feedback Prompting (AIFP) for Obstacle-Aware Path Planning via LLMs

📜 Overview

This repository contains the code and dataset for the paper:

"Adaptive Iterative Feedback Prompting for Obstacle-Aware Path Planning via LLMs."

📌 Abstract

Planning is a critical component for intelligent agents, especially in Human-Robot Interaction (HRI). Large Language Models (LLMs) demonstrate potential in planning but struggle with spatial reasoning. This work introduces Adaptive Iterative Feedback Prompting (AIFP), a novel framework that improves LLM-based path planning by incorporating real-time environmental feedback. AIFP prompts an LLM iteratively to generate partial trajectories, evaluates them for collision detection, and refines them when necessary using a Receding Horizon Planning (RHP) approach.

🔑 Key Features

✅ LLM-based path planning with adaptive feedback
✅ Collision-aware trajectory generation
✅ Iterative re-planning mechanism using Receding Horizon Planning (RHP)
✅ Handles static and dynamic obstacles
✅ Improves success rate by 33.3% compared to naive prompting
✅ Fully implemented with OpenAI's GPT-4 API

✅ A star and RRT planners in the same 2D domain

📂 Repository Structure

├── src/                   # Source code for AIFP framework
│   ├── aifp_planner.ipynb              # Core implementation of AIFP
│   ├── naive_llm_planner.ipynb         # A naive llm planner
│   ├── astar_planner.py                # A* Planner in the grid version of same 2D domain
│   ├── rrt_planner.py                  # RRT Planner in the same 2D domain
├── results/               # Outputs of path planning trials
├── README.md              # This README file
└── requirements.txt       # Required Python dependencies

🛠 Installation & Setup

1️⃣ Clone the repository:

git clone https://github.com/yourusername/AIFP-PathPlanning.git
cd AIFP-PathPlanning

📊 Experimental Results

Environment	AIFP Success Rate (%)	Naïve Prompting (%)
Single Obstacle	55.6%	22.3%
Double Obstacles	36.7%	14.0%
Random Obstacles	31.5%	12.5%
Moving Obstacle	48.5%	N/A
Moving Goal	51.5%	N/A

✔️ AIFP significantly outperforms naïve prompting, especially in static environments! 🚀

📌 Citation

If you use this work, please cite:

@article{AIFP2025,
  title={Adaptive Iterative Feedback Prompting for Obstacle-Aware Path Planning via LLMs},
  author={Masoud Jafaripour, Shadan Golestan, Shotaro Miwa, Yoshihiro Mitsuka, Osmar R. Zaiane},
  year={2025},
  Conference={AAAI LM4Planning Workshop}
}

🏗️ Future Work

• 🔹 Extend AIFP to 3D navigation tasks
• 🔹 Integrate Vision-Language Models (VLMs) for richer environmental perception
• 🔹 Explore graph-based path representations for improved trajectory optimization

---

🚀 Star ⭐ this repo if you find it useful!
📧 Feel free to submit issues, PRs, or suggestions.

this repo is being updating.

---

🖼️ Qualitative Planner Comparisons

The following visualizations compare LLM-based, learning-based, and classical planners under identical start–goal and obstacle configurations.
All figures are shown at a fixed size with compact captions for consistent visual comparison.

🤖 LLM-Based and Hybrid Planners

_{AIFP: Adaptive Iterative Feedback Prompting}

---

🧠 Learning-Based Planners

_{Q-learning: Handcrafted reward}

_{Q-learning: LLM-designed reward}

---

📐 Classical and Sampling-Based Planners

_{A*: Grid-based shortest path}

_{RRT: Sampling-based planner}

_{MCTS: Tree search planner}

---

🔍 Observations

• AIFP refines trajectories via iterative feedback and collision checking.
• LLM-informed reward learning biases learned paths toward smoother, goal-directed behavior.
• Classical planners offer strong geometric baselines but lack semantic adaptability.

All planners are evaluated in the same 2D environment with identical obstacle layouts.