FASTNav: Fine-tuned Adaptive Small-language-models Trained for Multi-point Robot Navigation

TL;DR

This work targets edge-deployed robot navigation by boosting small language models (SLMs) through LoRA-based fine-tuning and a teacher-student knowledge transfer framework. FASTNav combines three components—fine-tuning, teacher-student iteration, and a robot navigation controller using Navigation2—to transform natural-language directives into sequences of waypoints that a robot can execute locally. Empirical evaluation in simulation and real-world settings shows substantial improvements in success rate and responsiveness over other SLM-based and compressed setups, with performance approaching that of larger models while preserving edge efficiency and privacy. The approach demonstrates the practical potential of private, low-cost, language-guided navigation on edge devices, while identifying future work in multimodal integration and few-shot adaptation for broader applicability.

Abstract

With the rapid development of large language models (LLM), robots are starting to enjoy the benefits of new interaction methods that large language models bring. Because edge computing fulfills the needs for rapid response, privacy, and network autonomy, we believe it facilitates the extensive deployment of large models for robot navigation across various industries. To enable local deployment of language models on edge devices, we adopt some model boosting methods. In this paper, we propose FASTNav - a method for boosting lightweight LLMs, also known as small language models (SLMs), for robot navigation. The proposed method contains three modules: fine-tuning, teacher-student iteration, and language-based multi-point robot navigation. We train and evaluate models with FASTNav in both simulation and real robots, proving that we can deploy them with low cost, high accuracy and low response time. Compared to other model compression methods, FASTNav shows potential in the local deployment of language models and tends to be a promising solution for language-guided robot navigation on edge devices.

Figures (8)

• Figure 1 : FASTNav is a method for boosting SLMs for robot navigation. We improve the performance of SLMs for multi-point robot navigation by fine-tuning and teacher-student iteration.
• Figure 2 : Architecture of FASTNav. The architecture contains three components: (1) fine-tuning of SLMs based on PEFT: we use PEFT to give SLMs access to the necessary environment information and to constrain their output format; (2) teacher-student iteration: the fine-tuned SLMs cooperate with teacher models to solve navigation tasks, where the teacher serves as a prompt engineer and feedback receiver, transferring knowledge and helping SLMs correct mistakes and (3) robot navigation: the robot receives an ordered list of coordinate points from SLMs and moves with the navigation algorithms according to the list.
• Figure 3 : Human-in-the-loop generation of the dataset. The human-in-the-loop generation is used to generate our dataset for fine-tuning, which is a valid method for obtaining datasets aligned with humans. The yellow box corresponds to the first cycle, and the green box corresponds to the second cycle. $T_i$ means the i-th task while $S_i$ means the score that $T_i$ gets.
• Figure 4 : The accuracy curve during fine-tuning and iteration. This graph shows the accuracy of five models and GPT4 on the test set during fine-tuning and iteration processes. It can be seen that all the models' accuracy rises gradually in both processes and finally gets close to that of GPT4. (The last fine-tuning epoch is immediately followed by the first round of iterations, which is why they continue on the same axis.)
• Figure 5 : The sizes of the environments used in our experiments. The lab hallway and complex building are used in the real-world experiments, while the hospital is used in the simulation experiment.