Hybrid intelligence systems for reliable automation: advancing knowledge work and autonomous operations with scalable AI architectures

• Published in: Frontiers in robotics and AI Vol. 12; p. 1566623
• Main Authors: Grosvenor, Allan, Zemlyansky, Anton, Wahab, Abdul, Bohachov, Kyrylo, Dogan, Aras, Deighan, Dwyer
• Format: Journal Article
• Language: English
• Published: Switzerland Frontiers Media S.A 17.07.2025
• Subjects: CFD; computational fluid dynamics; hybrid intelligence (HI); joint embedding predictive architecture (JEPA); reinforcement learning (RL); Robotics and AI; space domain awareness; CFD; hybrid intelligence (HI); space domain awareness; joint embedding predictive architecture (JEPA); computational fluid dynamics; reinforcement learning (RL)
• ISSN: 2296-9144; 2296-9144
• DOI: 10.3389/frobt.2025.1566623

Mission-critical automation demands decision-making that is explainable, adaptive, and scalable-attributes elusive to purely symbolic or data-driven approaches. We introduce a hybrid intelligence (H-I) system that fuses symbolic reasoning with advanced machine learning a hierarchical architecture, inspired by cognitive frameworks like Global Workspace Theory (Baars, A Cognitive Theory of Consciousness, 1988). This architecture operates across three levels to achieve autonomous, end-to-end workflows: Navigation: Using Vision Transformers, and graph-based neural networks, the system navigates file systems, databases, and software interfaces with precision. Discrete Actions: Multi-framework automated machine learning (AutoML) trains agents to execute discrete decisions, augmented by Transformers and Joint Embedding Predictive Architectures (JEPA) (Assran et al., 2023, 15619-15629) for complex time-series analysis, such as anomaly detection. Planning: Reinforcement learning, world model-based reinforcement learning, and model predictive control orchestrate adaptive workflows tailored to user requests or live system demands. The system's capabilities are demonstrated in two mission-critical applications: Space Domain Awareness, Satellite Behavior Detection: A graph-based JEPA paired with multi-agent reinforcement learning enables near real-time anomaly detection across 15,000 on-orbit objects, delivering a precision-recall score of 0.98. Autonomously Driven Simulation Setup: The system autonomously configures Computational Fluid Dynamics (CFD) setups, with an AutoML-driven optimizer enhancing the meshing step-boosting boundary layer capture propagation (BL-CP) from 8% to 98% and cutting geometry failure rates from 88% to 2% on novel aircraft geometries. Scalability is a cornerstone, with the distributed training pipeline achieving linear scaling across 2,000 compute nodes for AI model training, while secure model aggregation incurs less than 4% latency in cross-domain settings. By blending symbolic precision with data-driven adaptability, this hybrid intelligence system offers a robust, transferable framework for automating complex knowledge work in domains like space operations and engineering simulations-and adjacent applications such as autonomous energy and industrial facility operations, paving the way for next-generation industrial AI systems.