2026/01/06 by Amar Salehi, Bairong Zhu, Haoyu Liu +1 · 1 voice
Computer Science · Physics and Astronomy · #Reinforcement Learning in Robotics #Micro and Nano Robotics #Adaptive Dynamic Programming Control
paper · pdf · doi:10.1002/aisy.202501053
openalex publication_date 2026/01/06 · openalex created_date 2026/01/08 · openalex updated_date 2026/07/23
Microrobots show great promise in biomedicine and environmental remediation, yet precise control in complex environments remains a significant challenge. Despite advancements in intelligent control systems, they often suffer from sample inefficiency and prolonged training times. This study presents a data‐driven framework to train deep reinforcement learning (DRL) algorithms for autonomous microrobot motion control. A supervised artificial neural network (SuANN) is trained to emulate microrobot–environment interactions based on data from a soft actor‐critic (SAC) model trained in a physical system. The truncated quantile critics (TQC) algorithm is then trained within this simulated environment. Integrated with A* path planning, the TQC‐SuANN model demonstrated superior real‐time obstacle avoidance and control accuracy in environments containing static and dynamic obstacles, as well as moving targets. Compared to the baseline SAC model, TQC‐SuANN achieved a 30.69% reduction in path deviation and a 23.43% increase in task completion speed. This approach significantly reduced training time, improved sample efficiency, and enhanced DRL performance for microrobot control. This framework enables scalable, efficient control of microrobots in complex environments.