This project investigates the robustness and generalization of perception models for autonomous driving under distributional shifts. The work consists of two complementary components. First, we benchmark state-of-the-art 3D multi-modal perception models, BEVFormer and BEVFusion, on the nuScenes-C dataset. Using PyTorch forward hooks, we extract high-dimensional latent feature representations to enable Out-of-Distribution (OOD) analysis through statistical evaluation of internal model activations. These models fuse multi-view camera and LiDAR data and contain deep, high-capacity architectures, making feature extraction computationally intensive and memory-demanding. Second, we study cross-domain generalization in semantic segmentation by evaluating DeepLabV3+, trained on Cityscapes, across the IDD and BDD100K datasets. The goal is to identify object classes that exhibit significant performance degradation when models are deployed across geographically distinct domains. This requires large-scale inference over high-resolution datasets and repeated forward passes to collect predictions for detailed class-wise analysis. Both components involve processing large datasets and running computationally expensive pipelines. The nuScenes-based experiments require handling multi-modal data exceeding 800 GB, while the segmentation benchmarks involve datasets on the order of hundreds of gigabytes. To ensure efficient and stable execution, the project requires access to high-performance computing resources, including GPUs with at least 24 GB VRAM (preferably 32–80 GB) to avoid out-of-memory errors during feature extraction, 1–2 TB of NVMe SSD storage to prevent I/O bottlenecks, and approximately 64 GB of system RAM with multi-core CPUs (16–32 cores) to support parallel data loading and preprocessing. Access to these resources is essential to enable reproducible, large-scale benchmarking and to support the development of robust perception systems for safety-critical autonomous driving applications.