Collaborative Aided Target Recognition Using Next-Generation Foundation Models for Multi-Domain Unmanned Systems

Modern military operations across air, land, and sea domains increasingly rely on heterogeneous teams of unmanned platforms tasked with detecting, identifying, and disambiguating targets at long range, often under appearance ambiguity, viewpoint and scale variation, sensor heterogeneity, and degraded Positioning, Navigation, and Timing (PNT). Today, reconciling observations across platforms is largely a manual, communication-intensive process that is slow, cognitively demanding, and error-prone. Single-platform AiTR cannot resolve perceptual aliasing (e.g., identical vehicle classes in coordinated formations) or fragmented observations caused by occlusion and partial visibility. Recent advances in foundation models offer transformative potential. Vision-Language Models (e.g., CLIP, LLaVA, Florence-2, Qwen-VL) provide view-invariant semantic representations; Vision-Language-Action models (e.g., RT-2, OpenVLA) couple perception with embodied decision-making; and State-Space Models (S4, S5, Mamba, Mamba-2) enable efficient long-sequence spatiotemporal reasoning with linear complexity—well-suited to multi-platform sensor fusion. However, these models are not yet adapted for collaborative, multi-platform military AiTR under realistic battlefield constraints. Offerors are encouraged to propose innovative technology-agnostic approaches built around next-generation foundation models (no legacy CNN-only solutions). Development should address: Collaborative Multi-Platform Target Disambiguation (foundation-model-based representations enabling cross-platform target correspondence under appearance ambiguity, viewpoint, scale variation, and clutter); Geometry-Consistent Multi-Modal Sensor Fusion (heterogeneous modalities such as EO/IR, SAR, LiDAR, RF across platforms with diverse resolution, FoV, and noise characteristics); Spatiotemporal World Modeling for Predictive AiTR (collective world models that maintain persistent target identity and forecast scene evolution under asynchronous, partial, intermittent data); Friend-Foe-Neutral (Gray) Classification (robust discrimination among blue, red, and gray entities); Resilient Operation in Degraded Environments (desirable capability to operate in GPS/PNT-denied, comms-degraded conditions); and an Agentic Autonomy Stack (a foundation-model-based framework orchestrating perception, reasoning, and tasking across heterogeneous platforms). Tri-Service Relevance: Army applications include multi-domain operations with ground/aerial scout teams; Navy/USMC applications include distributed maritime operations and expeditionary littoral reconnaissance; Air Force/Space Force applications include collaborative combat aircraft (CCA) and ISR drone swarms of varying size classes. Special Considerations: cybersecurity (model integrity, adversarial robustness against evasion/poisoning), supply-chain provenance of pre-training data, and explainability of model outputs shall be addressed.