Research on Multi-mode Path Planning Algorithm for Nanorobots Based on Improved Deep Reinforcement Learning

To address strong dynamic uncertainty and limited adaptability of traditional algorithms in nanorobot path planning, especially in microscopic textile fiber and porous-material environments, this study proposes a multi-mode path planning algorithm based on an improved deep Q-network (NP-IDQN). The method is applicable to nanorobots driven by fluid, chemical, or electromagnetic micro-actuation fields, where stable navigation must satisfy both trajectory efficiency and physical feasibility. The algorithm is optimized at three levels. At the network-architecture level, residual connections, layer normalization, and Transformer mechanisms are introduced to enhance feature extraction and global correlation modeling for fluid–obstacle interactions. At the training-strategy level, curriculum learning with multigeometry and multi-difficulty scenarios is adopted, covering rectangular, spherical, and mixed obstacles to improve environmental generalization. At the decision-making level, a multi-objective reward function integrates distance optimization, obstacle avoidance, energy efficiency, and microscopic driving-force constraints. Simulation results show that NP-IDQN improves trajectory stability and reduces the risk of local optima compared with conventional reinforcement learning and classical planning methods. The study provides technical support for precise navigation of nanorobots in complex fiber networks, functional material synthesis, environmental purification, and electromagneticactuated micro/nanorobotic systems.