EnRoute-CIL: Energy-Guided Routed Continual Learning for Intelligent Cockpit

This software release provides a reference implementation of ERO-MoE-CIL, a privacy-preserving open-world class-incremental learning framework for intelligent cockpit behavior recognition. The codebase integrates a frozen Vision Transformer backbone, replay with knowledge distillation, optional Mixture-of-Experts adapters, energy-based out-of-distribution scoring, and orthogonal gradient projection. Experiments are configured on Split CIFAR-100 (50 + 10x5) and evaluated with Average Accuracy (AA), Average Forgetting (AF), and optional OOD/runtime diagnostics (AUROC, FPR@95TPR, memory, and latency). This release is intended for research reproducibility and educational use.