Vision-Language-Action Soccer Robot—LLM Fine-Tuning and Dataset Creation
Led the construction and fine-tuning of a QLoRA-based language model targeted at on-device deployment for a robotics vision-language-action system. Created and augmented a 1,400-example dataset for fine-tuning, applying teacher-student distillation with LLaMA-3.1-8B to increase data volume and diversity. Conducted training and evaluation to optimize performance for real-time robotic actions involving structured inputs and outputs. • Developed and labeled deterministic datasets for LLM training focused on robotic task instructions. • Automated dataset augmentation using teacher-student model distillation. • Benchmarked and validated LLM performance via structured JSON-based I/O tasks for robotics control. • Orchestrated on-device integration and performance validation of the fine-tuned LLM.