Back Home

GitHub Repo

Autoware Announces 2.0: Adds an AI Planner Option; Model Directory Changes Require Redownloading

On November 2, the Autoware Foundation announced version 2.0, integrating learned planning into its existing autonomous driving software architecture and introducing model version management and a package index. The release still selects a single planner at startup; parallel generators and automatic trajectory selection are not yet part of the default startup flow.

:/ :\ · CC BY 2.0 · Image source
zh-Hant

The Autoware Foundation published its official 2.0 announcement on November 2, outlining updates to AI planning, model management, and system assembly. The 2.0.0 release had already appeared on GitHub on October 29. The update lets developers integrate learned planners into existing autonomous driving software and reproduce system configurations through versioned settings. Official announcement

The architecture centers on a generator–selector design: generators propose candidate trajectories, while selectors check, score, and choose among them. This points toward a shared trajectory interface for rule-based methods, diffusion models, and end-to-end approaches, reducing the effort needed to integrate new planning methods with existing control flows. However, the official documentation distinguishes the design from its implementation: Autoware 2.0.0 runs only one generator at a time, with rule_based as the default. Set planning_setting:=diffusion_planner to use the diffusion planner. Although the selector package exists, it is not yet loaded by the startup flow; currently, the trajectory modification module takes the first candidate and passes it to the planning validator. Architecture documentation

There is a direct compatibility change for deployment. Models are now managed using Hugging Face version tags, with Ansible’s artifacts role downloading pinned versions through the hf CLI. The new directory structure separates files by model variant, so files downloaded with version 1.9.0 or earlier cannot be used directly with the new launch files; models must be downloaded again when upgrading. The old setup-dev-env.sh has also been removed, so existing installation scripts need to be updated. Release notes

Another change is the Autoware Index, which registers sources and compatibility records for ROS 2 packages outside the main workspace. Its validation process records the tested commit, Autoware version, and build test results, while the CLI can write selected packages and their dependencies to a .repos list. This makes system assembly inputs and validation status easier to trace; these test records still do not replace evaluation in road scenarios. Index repository

The next things to watch are when multiple generators and the selector are actually connected to the execution flow, and how each combination performs in latency, failure handling, and scenario testing. For now, 2.0 can be treated as an integration starting point for evaluating AI planning; the official announcement provides no comparative data sufficient to show an overall improvement in driving safety.

Sources

  1. Autoware Foundation releases Autoware 2.0 to Expand AI Support for Open-Source Autonomous Driving
  2. Autoware 2.0.0 release notes
  3. Architecture 2.0
  4. Autoware Index repository