Fine-Tuning Small Language Models for Edge Robotics
researchMd. Shakhoyat Rahman ShujonOct 10, 2026

Fine-Tuning Small Language Models for Edge Robotics

Exploring LoRA adaptation, QLoRA, and 4-bit quantization benchmarks for real-time inference on resource-constrained embedded systems.

Md. Shakhoyat Rahman Shujon · admin

Abstract

Deploying large language models directly onto edge robotics platforms poses significant memory and power bottlenecks. In this paper, we explore parameter-efficient fine-tuning (PEFT) methodologies, specifically LoRA and quantized 4-bit NormalFloat (NF4) weights.

Methodology

We benchmarked Gemma-2B and Phi-3 Mini on Nvidia Jetson Orin Nano and Raspberry Pi 5 under varying thermal profiles.

Results

Quantization achieved a 3.8x memory reduction with less than 1.4% degradation in intent classification accuracy.