Manipulating heterogeneous microtargets based on optoelectronic tweezers (OETs) to construct micropatterns with specific distribution and ordered arrangement enables recapitulating the spatial architecture of cells in native tissues, and has significant potential in tissue regeneration, medical diagnostics, and cell behavior research. However, the uncertain disturbances in liquid environment, collision risk, and electrokinetic interference in OETs system can cause microtargets to deviate from the safe and controlled state, especially for manipulation tasks with heterogeneous microtargets. Here, we propose an improved hierarchical value iteration network (IHVIN)-GAT-based path planning method for parallel manipulation of heterogeneous microtargets with independent control, integrating goal assignment, feature extraction, and decentralized decision-making. The Kuhn–Munkres-based goal assignment model periodically modifies the matching relationship between microtargets and goal positions to reduce the task complexity. High-order features involving path planning are extracted by an IHVIN model, and then selectively aggregated and convolved through GAT to yield real-time locomotion strategies for all microtargets. For the issues of constraint variability and system heterogeneity, discrete locomotion constraints are developed through analysis of escape mechanism, then embedded into modeling procedures and converted to heterogeneous edge weights in graph domain. The simulation and experimental results demonstrate the desired performance of the IHVIN-GAT model in high timeliness, high strategy quality, and compatibility for microtarget number, where up to 20 microtargets from three categories are parallel manipulated within 16 s to form arbitrary micropatterns recapitulating microscale architecture of cells in native tissues. We anticipate that our method will contribute to construct more biomimetic microstructures with heterogeneous cells for biomedical applications in the future.