Abstract
UniAD organizes perception and planning as a unified autonomous driving stack. The paper argues that downstream planning can provide a useful objective for structuring upstream scene understanding.
Problem
Many driving systems optimize perception tasks independently, leaving the relationship between intermediate predictions and the final planning objective implicit.
Method
The system connects tracking, mapping, motion forecasting, occupancy prediction, and planning in a shared architecture. This creates a path for task-level information to influence the representations used by later decisions.
Personal insights
The most interesting idea is the choice of planning as the organizing principle. It gives the system a common direction without requiring every intermediate representation to be perfect in isolation.
For my research, the paper is a useful reference for asking which predictions are genuinely actionable and which are only convenient benchmarks.