Karl Iagnemma & Oscar Beijbom (Aptiv Autonomous Mobility) - MIT Self-Driving Cars

Key challenges include validating complex 'black box' systems like neural networks, ensuring the quality and representativeness of training data, verifying algorithm stability and invariance, and addressing the complexity of rare events and continuous code updates.

PointPillars is a novel lidar encoder that translates point clouds into a feature representation suitable for 2D convolutional networks, offering a significant speed-up compared to voxel-based methods like VoxelNet, by simplifying the feature extraction process.

NuScenes was created to provide a more comprehensive and up-to-date dataset for autonomous driving research, offering 1000 scenes with full 3D bounding box annotations, radar data, and diverse driving scenarios, advancing beyond the limitations of the older KITTI benchmark.

Aptiv uses a flexible 'rule book' system within its driving policy to adapt to different traffic rules, norms, and driving sides. They carefully design the system architecture to scale across diverse cities and countries, considering infrastructure, driving patterns, and regulatory environments.

It refers to embedding powerful but potentially less understood learning-based components (like trajectory proposers) within a robust safety architecture. This 'cage' or wrapper allows for rigorous claims about the overall system's performance and safety, even with complex internal models.

Perceived safety is crucial because even if a car is technically safe, riders must feel safe to build confidence and willingness to use the service. It's a subjective aspect that engineers must consider alongside technical safety metrics.

Models need to be trained on data representative of the deployment environment. While lidar point clouds are augmentable (e.g., simulating rare objects like pianos), deployment in conditions like snow requires specific training data capturing those scenarios.