Designing the Loss for a Multi-View Keypoint Detector
Heatmaps, soft-argmax, reprojection and multi-view consistency — and how the pieces fit into one objective.
Blog
2026
Tesla’s Autonomy Stack: Five Architectural Bets
Reading the public record — from per-camera detection to BEV fusion, occupancy, an end-to-end policy, and the world model that trains it.
Camera, Lidar, Radar: Localizing Without an HD Map
What each sensor actually measures, where each one degenerates, and how mapless stacks estimate pose and build map structure online.
Coordinate Conventions, or: Why Your Transform Is Inverted
Frames, rotations and transform chains across OpenCV, ROS 2 and GTSAM — and the notation that makes the bugs impossible.
Multi-Task Learning: One Backbone, Many Objectives
Why autonomous driving stacks share a backbone across tasks, and what makes the combined loss hard to balance.