A deep technical exploration of autonomous vehicle software architecture, focusing on optimization techniques across the full AV stack. Covers context-aware sensor fusion with dynamic covariance weighting for LiDAR, camera, and radar inputs based on driving scenario (highway vs. urban). Explains Model Predictive Control (MPC) formulation with cost functions, constraint handling, and solver choices (OSQP, Ipopt). Details real-time compute budget allocation across pipeline stages within a 100ms control cycle, deterministic scheduling with WCET budgets, and logging/explainability infrastructure using Protobuf schemas and MCAP formats. Includes pseudocode examples in Python and C++.
Table of contents
IntroductionThe End-to-End Architecture: From Sensor to ActuationOptimizing the Perception Pipeline: Dynamic Resource AllocationTrajectory Planning: The Mathematics of MPCReal-Time Compute Budget and MiddlewareDebugging and Explainability: The Data LayerFinal ThoughtsAbout the AuthorSort: