TriCLE tri-modal AI hits 88.33% aircraft clustering accuracy on edge

TriCLE, developed by Kishor Datta Gupta and six co-authors, tackles a hard real-world problem: classifying aircraft from aerial imagery on edge platforms with limited memory, compute, and intermittent connectivity. Rather than relying on calibrated thermal and LiDAR sensors that are often unavailable, TriCLE generates structure-preserving FLIR-style thermal views and pseudo-LiDAR depth projections from a single RGB image. These synthetic modalities are fused with task instructions in a compact Qwen3-VL backbone, aligned to an expert taxonomy based on propulsion, airframe family, size, design era, and configuration—so outputs reflect engineering similarity, not just surface appearance.

The team evaluated supervised fine-tuning, rotation-preserving SFT, and three policy-alignment strategies (GRPO, GSPO, DAPO). Sequence-level GSPO delivered the strongest results: 88.33% validation accuracy and 0.91 weighted F1, with 94.00% parseable output formatting on a held-out test set. After 4-bit quantization and attention-memory optimization, the aligned 4B model fits an 8GB deployment target and processes each tri-modal triplet in 1.48 seconds. This makes TriCLE a practical prototype for interpretable, edge-feasible aircraft grouping, though the authors stress the need for validation on real aligned thermal and LiDAR streams before production deployment.