Artificial intelligence

Abstract: A paper studies the ability of compact large language models (LLMs) to perform local path planning using only limited, sensor-like observations. While recent work typically employs LLMs as high-level planners within hybrid systems leaving low-level navigation to classic pathfinding algorithms, considerably less attention has been given to how smaller models behave when placed directly in the control loop. To address this gap, we evaluate a 1.7-billion-parameter model (Qwen3-1.7B) as a local planner in a continuous two-dimensional environment. The agent navigates toward a target within a bounded rectangular map containing varying numbers of obstacles. At each step, a ring of collision-free candidate waypoints is generated, and made available to the model through a textual table containing indices, coordinates, and distances to the goal. Given few-shot demonstrations, the model selects a single candidate index to determine the next move. We report performance in terms of success rate, path length, and efficiency relative to the straight-line distance, and analyze the effect of hyperparameters tuned via Optuna. The results provide the first quantitative characterization of Qwen3-1.7B as a local path planner, highlighting both its capabilities and limitations. The findings motivate further exploration of richer observation formats, memory mechanisms, and comparisons with algorithmic baselines.

Keywords: large language models, pathfinding, navigation, local planning, machine learning.

References:

1. Huang, W., Abbeel, P., Pathak, D., & Mordatch, I. (2022). Language Models as Zero-Shot Planners: Extracting Actionable Knowledge for Embodied Agents. ArXiv (Cornell University). https://doi.org/10.48550/arxiv.2201.07207
2. Ahn, M., Brohan, A., Brown, N. A., Yevgen Chebotar, Andres, O., David, B., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Daniel, Hsu, J., Ibarz, J., Ichter, B., Irpan, A., Jang, E. B., Ruano, R. J., Jeffrey, K., Jesmonth, S., & Joshi, N. J. (2022). Do As I Can, Not As I Say: Grounding Language in Robotic Affordances. https://doi.org/10.48550/arxiv.2204.01691
3. Liang, J., Huang, W., Xia, F., Xu, P., Hausman, K., Ichter, B., Florence, P., & Zeng, A. (2022). Code as Policies: Language Model Programs for Embodied Control. ArXiv (Cornell University). https://doi.org/10.48550/arxiv.2209.07753
4. Doma, P., Arab, A., & Xiao, X. (2024, December 3). LLM-Enhanced Path Planning: Safe and Efficient Autonomous Navigation with Instructional Inputs. https://doi.org/10.48550/arXiv.2412.02655
5. Meng, S., Wang, Y., Yang, C.-F., Peng, N., & Chang, K.-W. (2024). LLM-A*: Large Language Model Enhanced Incremental Heuristic Search on Path Planning. ArXiv (Cornell University). https://doi.org/10.48550/arxiv.2407.02511
6. Chen, J., Lin, B., Xu, R., Chai, Z., Liang, X., & Wong, K.-Y. (2024). MapGPT: Map-Guided Prompting with Adaptive Path Planning for Vision-and-Language Navigation. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 9796–9810. https://doi.org/10.18653/v1/2024.acl-long.529
7. Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., Zheng, C., Liu, D., Zhou, F., Huang, F., Hu, F., Ge, H., Wei, H., Lin, H., Tang, J., & Yang, J. (2025). Qwen3 Technical Report. ArXiv (Cornell University). https://doi.org/10.48550/arxiv.2505.09388
8. Watanabe, S. (2023, May 26). Tree-Structured Parzen Estimator: Understanding Its Algorithm Components and Their Roles for Better Empirical Performance. ArXiv.org. https://doi.org/10.48550/arXiv.2304.11127