The rapid integration of robotics and artificial intelligence (AI) has accelerated the development of autonomous systems; however, existing robot automation still relies heavily on traditional AI and policy-based control, which suffers from limited co...
The rapid integration of robotics and artificial intelligence (AI) has accelerated the development of autonomous systems; however, existing robot automation still relies heavily on traditional AI and policy-based control, which suffers from limited contextual perception, weak generalization to high-level instructions, and high re-engineering costs when environments change. Recent Large Language Model (LLM)-based approaches have shown promise in addressing these challenges but remain impractical for real-world robotic control due to excessive model size, high latency, limited long-term reasoning, and substantial hardware requirements.
To overcome these limitations, this research proposes a Lightweight Robot Control System that integrates an on-device Small Language Model (sLM) with Digital Twin (DT) technology. The proposed system introduces a resource-efficient Robot sLM optimized for perception, planning, and control, in contrast to human-centric LLMs. A parameter-efficient fine-tuning strategy using LoRA adapters enhances perception accuracy while reducing latency, enabling on-device operation with minimal computational overhead. In addition, Planning and Perception Adapters are designed to enable long-term multi-skill task execution by decomposing high-level goals into step-level actions, maintaining memory, and performing iterative perception–planning–action loops.
A DT-based simulation environment is developed to validate robot behaviors prior to real-world execution, allowing predictive testing, failure diagnosis, and safe policy refinement. The integrated framework demonstrates its effectiveness across multiple scenarios—including multi-room search, maze escape, and target following—using only visual input. Experimental results show improved perception accuracy, inference latency reduction, and generalized multi-skill execution across unseen environments, confirming its practicality for scalable robotic deployment.
By eliminating cloud dependency and enabling low-latency, and scalable intelligence on physical robots, this research advances a practical automation framework that supports adaptive, long-term, and multi-skill robot control across complex environments.