A recent study published in Engineering unfolds a novel deep learning (DL)-powered framework tailored for the environment-adaptive navigation of size-adaptable microswarms. Spearheaded by Lidong Yang and Li Zhang, this research is set to overcome the long-standing challenges that have plagued the development of microswarm navigation.
Micro/nanorobots have emerged as a revolutionary force in biomedicine. As the paper details, they hold immense potential in a plethora of applications, including minimally invasive surgery, targeted delivery, and therapy. For instance, previous studies have reported successful targeted delivery using drug-loaded biohybrid microrobots, highlighting their significance. However, the journey towards autonomous collision-free navigation of microswarms in confined spaces has been fraught with difficulties. Microswarms vary in dimensions, and conventional path-planning methods often overlook the physical dimensions of microrobots. Moreover, they struggle to manage the reconfigurability of deformable microrobots, leaving a crucial gap in the field.
To bridge this gap, the research team engineered a cutting-edge DL-based scheme. At its core lies a deep Q-network (DQN) designed for path planning. This reinforcement learning - assisted approach excels in online planning, making it ideal for complex scenarios such as channel-like environments and those cluttered with dynamic obstacles. The DQN model underwent rigorous training on four distinct sizes of microrobots, specifically 1×1, 2×2, 3×3, and 5×5 pixels. Impressively, after 30,000 training episodes, the search success rate soared beyond 92%. This not only demonstrates the model's effectiveness but also its ability to provide an adaptive safe distance for diverse microrobots, ensuring reliable navigation. In experiments involving a microbead and a microswarm, the DQN-based planning strategy enabled both agents to deftly avoid obstacles and reach their intended targets, even in the face of dynamic obstacles.
Beyond motion planning, the researchers introduced a pattern-distribution planner. They zeroed in on real-time swarm pattern planning and control, devising a cost function to identify the optimal swarm distribution and a deep convolutional neural network (DCNN)-based system. Taking the ribbon-like swarm as a prime example, the DCNN could generate an optimal swarm pattern distribution by factoring in the surrounding obstacles. Whether in open spaces, around obstacles, or navigating through channels, the proposed framework could fine-tune the swarm's shape ratio and pointing angle, guaranteeing safe and efficient navigation.
This novel DL-based framework outshines conventional planning methods in dynamic scenarios and channel environments. It also showcases remarkable adaptability to microrobots of various sizes, offering a more robust and efficient navigation solution. Looking ahead, this study serves as a springboard for further exploration of microswarms' autonomy in more complex and practical working environments, such as physiologic settings or those with hydrodynamic disturbances. It has the potential to revolutionize the microswarm-related applications in biomedicine and other fields, marking a significant milestone in the scientific community.
The paper "A Deep Learning-based Framework for Environment-adaptive Navigation of Size-adaptable Microswarms," authored by Jialin Jiang, Lidong Yang, Shihao Yang, Li Zhang. Full text of the open access paper: https://doi.org/10.1016/j.eng.2024.11.020
