Hierarchical DRL Planning for Manned Unmanned Platforms

The patent segments the DRL training process into multiple parallel environments, where multiple instances of the same environment are created and trained simultaneously. This allows the system to explore diverse scenarios and improve generalization to new environments while reducing the total training time through parallel processing. The segmented approach enables efficient utilization of computational resources across multiple environment instances.

The patent introduces a hierarchical dimension to the DRL architecture, organizing the system into multiple layers: a global planner that handles high-level team-level strategies and semantic goal definition, and local planners that handle specific platform actions. This dimensional organization separates complex strategic reasoning from tactical execution, reducing the computational burden on individual components while maintaining overall system intelligence and automation capability.

The patent implements a universal environment configuration system where a single base environment class can be instantiated multiple times with different parameters and settings. This multi-functional design allows the same environment framework to serve multiple training purposes simultaneously, managing diverse training scenarios through a unified interface. The universal configuration approach reduces management complexity by providing consistent control mechanisms across all parallel environments while maintaining training efficiency through simultaneous execution.