A Robust Deep Ensemble Architecture for Safety‐Critical Decision‐Making in Autonomous Systems

ABSTRACT Real‐world, unpredictable conditions for autonomous vehicles continue to be a challenge. This study presents a Python‐developed “Redundancy Net” as a multi‐model ensemble method in improving the autonomous vehicle systems in perception, decision making, fault tolerance, adaptability, and robustness. Evaluated with the KITTI Vision Benchmark Suite, the framework was developed through the integration of stereoscopic cameras, LiDAR, GPS, and IMU data fusion with the implementation of hardware sensor fusion. For analysis, redundancy net data pre‐processing sequence normalization, voxelization, Kalman filtering, and sequence temporal alignment, the multimodal systems. Redundancy net uses the complementary features of CNNs for spatial perception, LSTMs for sequence learning, and transformers for global spatio‐temporal integration. In the redundancy and reliability layer, model predictions are dynamically fused with “confidence” based weighting through sensor faulted downgrades ensuring dependable outcomes. A fail‐safe decision module with Monte Carlo dropout and entropy‐based uncertainty evaluates system low confidence outcomes and activates safety holds. The Online Validation and Self‐Adaptation mechanism improves and enhances Self‐Adaptation mechanisms in real time by adjusting a model's parameters based on continuous performance evaluations and feedback. The experimental evaluations show that Redundancy Net provides a performance boost of 94.8% accuracy as compared to standalone CNN, LSTM, and Transformer models, which speaks to the framework's contribution to developing safe, flexible, and robust autonomous navigation in complicated driving environments.