Semantic Diverse Redundancy for Robust Multi-Modal Safety-critical AI Tasks

Dual Modular Redundancy (DMR) and Triple Modular Redundancy (TMR), often combined with diversity techniques, are widely used in safety-critical systems to achieve fault detection and/or tolerance. Traditional redundancy expects bit-identical outputs, flagging any mismatch as an error. However, emerging AI-based functionalities, such as camera and LiDAR-based object detection, are intrinsically stochastic and require only semantic correctness, allowing some variation in the outputs (e.g., slightly different confidence scores). In this work, we extend our previous semantic redundancy approach – originally developed for camera-based object detection – to LiDAR-based systems, which operate on 3D point cloud data. We propose software-only DMR and TMR schemes that introduce data-level diversity through domain-specific input transformations, preserving semantic meaning while increasing robustness. Our findings demonstrate that the method can be generalized to different AI tasks, helping improve safety and reliability in the different AI components of safety-critical systems.