Digital twin and AI-driven robotic embodied control system: a novel adaptive learning and decision optimization method

Abstract

Traditional robot control methods often encounter limitations such as lengthy development cycles and insufficient flexibility when addressing dynamic production environments and complex task requirements. To overcome these challenges, this paper constructs an integrated robot embodied control (EC) system that organically combines digital twins (DT), machine vision, and deep reinforcement learning (DRL). The method follows a closed-loop perception-decision-action framework. First, machine vision senses the environment in real time and precisely maps the 3D pose of the target object to the DT space. Second, DRL is conducted in the DT environment for training and strategy optimization. Finally, continuous state synchronization between the physical robot and the DT enables cross-environment policy transfer and online optimization. Taking the robotic arm pressing an emergency stop button as a representative task scenario, experimental results show that the system achieves a task success rate of 88% in the DT environment and 73% in the real physical environment, which was further improved to 76% through fine-tuning. In an extended lamp switch task, the success rate reached 79%, further verifying the generality and cross-environment adaptability of the framework. Overall, this integrated system significantly enhances the intelligence and operational efficiency of robotic systems, demonstrating its potential for achieving programming-free autonomous control in complex industrial environments.