Skip to content

Autoware 2.0 Architecture#

Warning

Under Construction

Generator - Selector architecture: conceptual overview#

generator-selector-simple.svg

Traditional autonomous driving follows a fixed pipeline:

Sensing → Perception → Localization → Planning → Trajectory

This works well for rule-based planners, but newer approaches like E2E or diffusion models don’t fit neatly into that structure. They may skip or replace parts of the pipeline, making integration difficult.

To support both classical and modern approaches, we abstracted away the front half of the pipeline.

Generators: Flexible trajectory producers#

A Generator is any module that outputs trajectories. It could be one or more:

  • rule-based or optimization planners using perception and maps
  • E2E models using raw sensor input
  • learned or sampling-based planners

Generators can reuse Autoware’s sensing, perception, localization, and control. Or bypass them. Multiple generators can run in parallel.

Selector: Safety + final choice#

The Selector receives candidate trajectories and:

  • Safety-checks them (e.g., rule compliance, drivable area)
  • Ranks and selects the best one based on context or driving policies

This enables:

  • Seamless integration of both robotics-based and E2E planners
  • Safe use of black-box models through explicit checks
  • Flexible experimentation with new planning methods
  • Robust decision-making by comparing multiple trajectory proposals